WO2021031704A1 - 对象追踪方法、装置、计算机设备和存储介质 - Google Patents
对象追踪方法、装置、计算机设备和存储介质 Download PDFInfo
- Publication number
- WO2021031704A1 WO2021031704A1 PCT/CN2020/099170 CN2020099170W WO2021031704A1 WO 2021031704 A1 WO2021031704 A1 WO 2021031704A1 CN 2020099170 W CN2020099170 W CN 2020099170W WO 2021031704 A1 WO2021031704 A1 WO 2021031704A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- current frame
- candidate
- matching
- feature
- target object
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
Definitions
- This application relates to the field of artificial intelligence technology, in particular to the field of image processing technology, and in particular to an object tracking method, device, computer equipment and storage medium.
- the inventor realizes that in the traditional method, the current frame is matched with the previous frame, so as to determine the target object to be tracked from the current frame according to the matching difference. In this way, if the target object in the previous frame is blurred or partially occluded, it will result in less effective feature information in the previous frame, resulting in inaccurate matching and inaccurate object tracking.
- an object tracking method, device, computer equipment, and storage medium are provided.
- An object tracking method including:
- the target object is identified from the candidate objects in the current frame.
- An object tracking device includes:
- the feature extraction module is used to select the current frame from a multi-frame image; perform feature extraction on the target object in the previous frame of the current frame to obtain the first feature of the target object; In each frame before the frame, extract the feature of the target object respectively to obtain the second feature of the target object; extract the feature of each candidate object included in the current frame;
- the feature matching module is configured to match the feature of each candidate object with the first feature to obtain a first matching result, and match the feature of each candidate object with the second feature to obtain a second matching result; The first matching result and the second matching result corresponding to the same candidate object, determining the final matching result between each candidate object and the target object; and
- the object recognition module is used to identify the target object from the candidate objects in the current frame according to the final matching result.
- a computer device includes a memory and one or more processors.
- the memory stores a computer-readable storage medium.
- the one or more processors execute the various embodiments of the present application. Steps in the object tracking method.
- One or more computer-readable storage media storing computer-readable instructions.
- the one or more processors execute the object tracking methods in the embodiments of the present application. step.
- the above object tracking method, device, computer equipment and storage medium respectively acquire the first feature of the target object in the previous frame and the second feature of the target object in the frames before the current frame, which is equivalent to considering that the target object is in The first feature in the closer previous frame also takes into account the second feature of the target object in the previous multiple frames, so that the extracted target feature has more information and is more accurate.
- the feature of each candidate object in the current frame is matched with the first feature and each second feature of the target object, and the final matching result between the candidate object and the target object is determined according to the first matching result and the second matching result,
- the matching result between each candidate object and the target object can be made more accurate, and the target object can be more accurately identified from the current frame based on the matching result, and the target object can be tracked more accurately.
- Fig. 1 is an application scenario diagram of an object tracking method according to one or more embodiments
- Fig. 2 is a schematic flowchart of an object tracking method according to one or more embodiments
- FIG. 3 is a schematic diagram of the principle of an object tracking method according to one or more embodiments.
- Fig. 4 is a block diagram of an object tracking device according to one or more embodiments.
- Figure 5 is a block diagram of a tracking and matching module according to one or more embodiments.
- Figure 6 is a block diagram of a computer device according to one or more embodiments.
- Figure 7 is a block diagram of a computer device in another embodiment.
- the object tracking method provided in this application can be applied to the application environment as shown in FIG. 1.
- the terminal 110 communicates with the server 120 through the network through the network.
- the terminal 110 may be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices.
- the server 120 may be implemented as an independent server or a server cluster composed of multiple servers. It can be understood that the terminal 110 has an image collection function.
- the terminal 110 may collect video and send the video to the server 120, and the server 120 executes the object tracking method in each embodiment of the present application according to the video.
- the server 120 may select the current frame from the multi-frame images of the video, and perform feature extraction on the target object in the previous frame of the current frame to obtain the first feature of the target object; from the multi-frame image located in the current frame In the previous frames, extract the features of the target object to obtain the second feature of the target object; extract the feature of each candidate object included in the current frame; compare the feature of each candidate object with the first feature The feature is matched to obtain a first matching result, and the feature of each candidate object is matched with the second feature to obtain a second matching result; according to the first matching result and the second matching result corresponding to the same candidate object, Determine the final matching result between each candidate object and the target object; according to the final matching result, identify the target object from each candidate object in the current frame. Further, the server 120 may add a highlight mark to the identified target object in the current frame, and notify the terminal 110
- the terminal 110 itself may also execute the object tracking method in each embodiment of the present application on each frame of image, without sending to the server 120. There is no restriction on who performs the object tracking method.
- an object tracking method is provided.
- the method is applied to a computer device as an example for description.
- the computer device may be the server 120 in FIG. 1, and includes the following steps:
- the current frame is the currently processed image frame.
- the multi-frame images may be multiple images in a video (for example, surveillance video).
- the object tracking method may not be limited to the scene where the object is tracked in the video, and may be applicable to the scene where the object is tracked in any multi-frame image. Therefore, in other embodiments, multiple frames of images may also be multiple pictures.
- S204 Perform feature extraction on the target object in the previous frame of the current frame to obtain the first feature of the target object.
- the previous frame of the current frame is the previous image frame adjacent to the current frame in the multi-frame image, that is, the latest frame before the current frame in the multi-frame image.
- Target object refers to the object that needs to be tracked.
- the target object exists in the previous frame, and the target object needs to be identified in the current frame, so as to realize the tracking of the target object from the previous frame to the current frame.
- the first feature of the target object is the feature of the target object in the previous frame of the current frame.
- steps S204 to S206 and S210 to S214 may be performed for each target object, so as to track each target object in the current frame.
- the computer device can identify the target object from the previous frame of the current frame, and extract the characteristics of the target object through at least one of detection processing such as convolution processing or edge, bone structure, and color detection to obtain the target object The first feature. It can be understood that since the first feature of the target object is extracted from the previous frame closest to the current frame, the first feature is equivalent to the recent feature of the target object.
- S206 Extract the feature of the target object from each frame before the current frame in the multi-frame image to obtain the second feature of the target object.
- Each frame before the current frame in the multi-frame image refers to the image frame before the current frame in the multi-frame image. It can be understood that when the multi-frame images are multiple image frames in a video frame, each frame located before the current frame is a video image frame generated before the current frame in the multi-frame image.
- the computer device may separately detect the target object in each frame before the current frame in the multi-frame image, and extract the characteristics of the target object from them, and obtain the second characteristics of the target object corresponding to each frame.
- the computer device can extract the features of the target object through convolution processing or detection processing such as edge, bone structure, and color detection for each frame located before the current frame to obtain each second feature of the target object.
- the processing is performed based on the second feature of the target object extracted from the frames before the current frame in the multi-frame image.
- the second feature of the target object is naturally not extracted. Not under consideration.
- the second feature of the target object is extracted from the frames before the current frame in the multi-frame image. Since the object tracking method has been executed in the frames before the current frame, it is relative to The current frame belongs to the historical frame (that is, the image frame for which the object tracking method in each embodiment of the present application has been executed). Therefore, the second feature of the target object extracted from each frame before the current frame, Equivalent to the historical characteristics of the target object.
- the feature of the target object may be extracted from all or a part (ie at least part) of the image frame located before the current frame in the multi-frame image to obtain the second feature of the target object.
- the computer device may select a preset number of frames before the current frame from the multi-frame image, and extract the features of the target object from the selected frames to obtain the second feature of the target object. It can be understood that the feature of the target object is extracted for each selected frame, so that the second feature of the target object corresponding to each selected frame can be obtained.
- the computer device may also select a preset number of frames in order from the current frame.
- the current frame is the 5th image frame in the video
- the preset number is 3
- the selected historical frame is 3 image frames from the 5th image frame forward, that is, the second to fourth images
- the frame is the selected image frame
- the features of the target object are extracted from the selected second to fourth image frames, respectively, to obtain the second feature of the target object corresponding to the second to fourth image frames respectively.
- the selected frame includes the frame before the current frame. It can be understood that when the preset number is greater than 1, the previous preset number of historical frames include not only the previous frame of the current frame, but also the frame before the previous frame.
- the computer device may randomly select frames that meet the preset number among the frames located in the current frame. For example, if the current frame is the fifth image frame in the video, and the preset number is 3, you can randomly select the first, third, and fourth image frames, and then select the first, third, and fourth images from In the frames, the features of the target object are extracted respectively to obtain the second features of the target object corresponding to the first, third, and fourth image frames.
- Candidate objects refer to objects that can be observed in the current frame and used to be judged as target objects. There is at least one candidate.
- the candidate object may include at least one of people, vehicles, animals, and objects.
- the computer device may recognize all objects in the current frame as candidate objects. For example, if there are two people, a dog, and a car in the current frame, the computer device can directly use the objects in the current frame as candidates regardless of the object category, that is, the two people, a dog, and a car are all candidates. As a candidate.
- the computer device may also select some objects as candidate objects from the objects that can be observed in the current frame. Specifically, the computer device can obtain the category of the target object, and from the objects that can be observed in the current frame, select an object that matches the category of the target object as a candidate object. For example, there are two people, a dog, and a car in the current frame, and assuming that the target object category is "people", then two "people" in the current frame can be selected as candidate objects.
- the computer device can perform edge detection (shape dimension), bone structure detection (internal structure dimension), and color detection (color dimension) for each candidate object in at least one dimension, so as to obtain the characteristics of each candidate object.
- edge detection shape dimension
- bone structure detection internal structure dimension
- color detection color dimension
- the computer device may perform convolution processing on each candidate object, and extract the characteristics of the candidate object through multiple rounds of convolution processing.
- S210 Match the feature of each candidate object with the first feature to obtain a first matching result, and match the feature of each candidate object with the second feature to obtain a second matching result.
- the computer device can match the feature of each candidate object with the first feature of the target object to obtain the first matching result corresponding to each candidate object.
- the computer device can match the feature of each candidate object with the second feature of the target object to obtain a second matching result corresponding to each candidate object.
- the first matching result may include matching success and matching failure.
- the second matching result may also include matching success and matching failure. That is, the matching result may not be in a numerical form, but a direct conclusion description form.
- the first matching result includes the first matching degree between the feature of the candidate object and the first feature of the target object.
- the second matching result includes the second degree of matching between the feature of the candidate object and the second feature of the target object.
- the first degree of matching may be characterized according to the first difference between the feature of the candidate object and the first feature of the target object. That is, the first difference value is used to characterize the feature matching degree between the feature of each candidate object and the first feature of the target object.
- the second matching degree may be characterized according to a second difference value between the feature of the candidate object and the second feature of the target object. That is, the second difference value is used to characterize the feature matching degree between the feature of each candidate object and each second feature of the target object.
- the corresponding second features can be extracted from the multiple frames, that is, there are multiple second features of the target object. Then, the feature of each candidate object and each second feature of the target object have a second difference value, that is, multiple second difference values can be obtained.
- the computer device may obtain the Euclidean distance between the feature of the candidate object and the first feature of the target object to obtain the first difference value, and obtain the candidate object's The Euclidean distance between the feature and each second feature of the target object obtains the second difference value between the feature of the candidate object and each second feature of the target object.
- S212 Determine a final matching result between each candidate object and the target object according to the first matching result and the second matching result corresponding to the same candidate object.
- the final matching result is that the candidate object is the target object.
- the computer device may also determine the final feature matching degree corresponding to each candidate object according to the first matching degree and the second matching degree corresponding to the same candidate object.
- the final feature matching degree is the final matching result. It can be understood that when the first degree of matching and the second degree of matching are characterized by the first difference value and the second difference value, the computer device may also determine according to the first difference value and the second difference value corresponding to the same candidate object The matching difference between each candidate object and the target object. It can be understood that the matching difference between each candidate object and the target object is the final matching result between the candidate object and the target object.
- the candidate object is identified from the current frame to obtain the target object.
- the computer device can identify the target object from the candidate objects in the current frame according to the difference in matching between each candidate object and the target object. For example, from the candidate objects in the current frame, the candidate object with the smallest matching difference value is selected to obtain the target object.
- the computer device may also highlight the selected object in the current frame to reflect that the object is the target object to be tracked.
- the above object tracking method separately obtains the first feature of the target object in the previous frame and the second feature of the target object in the frames before the current frame, which is equivalent to considering both the recent features of the target object and the target object
- the historical features of the extracted target features have more information and more accuracy.
- the feature of each candidate object in the current frame is matched with the first feature and each second feature of the target object, and the final matching result between the candidate object and the target object is determined according to the first matching result and the second matching result, As a result, the matching result between each candidate object and the target object can be made more accurate, and the target object can be more accurately identified from the current frame based on the matching result, and the target object can be tracked more accurately.
- step S208 extracting features of each candidate object in the current frame includes: obtaining at least one preset object detection template; respectively matching each object included in the current frame with the object detection template; determining the successfully matched object As a candidate.
- the object detection template is a preset template for detecting objects.
- the computer device can preset an object detection template.
- the computer device can match the object in the current frame with the object detection template, and use the object in the current frame that matches the object detection template as candidate objects. Furthermore, the feature of the candidate object is extracted from the current frame.
- a corresponding object detection template can be set with the category of the object as the dimension. That is, one type of object detection template is used to detect one type of object.
- the object detection template can be a human body frame template.
- the computer device can detect the frame of the object in the current frame and match the frame with the human body frame template. If the matching is successful, the object is a human body. , So as to detect and recognize all people in the current frame. Then, all the people identified are candidates.
- the object detection template can be a vehicle frame template (that is, a template that can identify the characteristics of the vehicle).
- the computer device can match the object in the current frame with the template. If the matching is successful, Explain that the object is a vehicle, thereby identifying all vehicles in the current frame. Then, all the vehicles identified are candidates.
- a corresponding object detection template can also be set according to the secondary classification.
- the secondary classification can be made into “old man”, “child”, “man” and “ woman”, etc.
- the object detection template can be set corresponding to the secondary classification, for example, child Detect the template, then all children can be identified from the current frame. In this way, it can be used to automatically find missing children.
- the computer device may obtain the category of the target object, obtain the object detection template corresponding to the category of the target object, and then match the object in the current frame with the object detection template, and compare the object detection template in the current frame with the object detection template.
- the matched objects are used as candidates.
- an object detection template for detecting "person”, such as a human body frame template, is obtained, so that the person in the current frame is detected as a candidate object.
- the objects in the current frame can be filtered through the object detection template to filter out the objects that meet the requirements as candidate objects. In this way, there is no need to perform differential matching of all objects in the current frame with the target object. , Thus saving computing resources while ensuring the accuracy of object tracking.
- the target object belongs to a pedestrian object;
- the object detection template includes a human body frame template.
- respectively matching each object included in the current frame with the object detection template includes: separately matching each object included in the current frame with the human body frame template.
- Determining the successfully matched object as the candidate object includes: identifying the pedestrian object included in the current frame as the candidate object according to the matching result.
- Pedestrian objects are the humanoid image content displayed in the image frame.
- the human body frame template is a pre-set human body frame. It can be understood that the human body frame contains the common characteristics of the human body. That is, except for special people, normal people can usually match the human body frame.
- the computer device may first detect the objects included in the current frame, and then match each object with the human body frame template, and according to the matching result, identify the pedestrian objects included in the current frame as candidate objects. It can be understood that when the matching result matches the body frame template of the object, the object is identified as a pedestrian object, and the object is a candidate object.
- the human body frame template can be set from the perspective of outline, that is, the human body frame template includes a common human body contour, and the computer device can match the outline of each object in the current frame with the human body frame template.
- the respectively matching each object included in the current frame with the human body frame template includes: for each object included in the current frame, performing edge recognition on the object to obtain the External contour feature; matching the external contour feature of the object with the human body frame template; when the matching is successful, it is determined that the object is a pedestrian object.
- the external contour feature is the feature of the external contour of the object and is used to characterize the external contour of the object.
- the computer device may detect the edge feature points of each object included in the current frame through edge detection processing, and connect the edge feature points in order to obtain the outer contour line, which is the outer contour characteristic data.
- the computer device can match the obtained external contour feature data of each object with the human body frame template including the common contour of the human body. When the matching is successful, it is determined that the object is a pedestrian object.
- the human body frame template can also be set from the perspective of human bones, that is, the human body frame template is a human bone frame template.
- the human skeleton framework template includes preset key points of appearance skeleton.
- the key points of the appearance bone are the bone points that are exposed to the outside and can be directly seen from the appearance.
- the computer device can match each object in the current frame with the human skeleton frame template including the preset appearance bone key points, and when it matches, it is determined that the object is a pedestrian object, that is, a candidate object.
- the target object belongs to a pedestrian object; the current frame includes multiple objects.
- the method further includes: for each object in the current frame, intercepting the object area map including the object from the current frame; the object occupies the main area of the object area map; respectively inputting each object area map to the pre-training In the human body recognition model of, the recognition results for the objects included in each object area map are output; when the recognition results characterize the object as a human body, the object is determined as a candidate object.
- the object area map is an image where the object occupies the main area. It can be understood that the object area map is a part of the interception from the current frame. For example, if the current frame includes three objects A to C, and the object area map of A is intercepted from them, then in the object area map of A, A occupies the main area.
- the human body recognition model is a pre-trained machine learning model used to recognize people.
- the human body recognition model can be obtained by iterative machine learning training through sample data in advance.
- the sample data includes sample images and object category labels.
- the image content in the positive sample image includes people, and the image content in the negative sample image includes objects other than people.
- the object category corresponding to the positive sample image is labeled as a human body label, and the object category corresponding to the positive sample image is labeled as a non-human body label.
- the computer device can input each object area map into a pre-trained human body recognition model, and output the recognition result of the object corresponding to each object area map; when the recognition result indicates that the object is a human body, the object is determined as a candidate object. When the recognition result indicates that the object is a non-human body, it is determined that the object is not a candidate object.
- the first matching result includes a first matching degree between the object and the target object; the second matching result includes a second matching degree between the object and the target object.
- step S212 includes: determining the final matching degree between each object and the target object according to the first matching degree and the second matching degree corresponding to the same object.
- Step S214 includes: selecting the smallest final matching degree from the final matching degrees corresponding to each object; selecting the object corresponding to the smallest final matching degree from the objects in the current frame to obtain the target object.
- the first degree of matching may be characterized according to the first difference between the feature of the candidate object and the first feature of the target object.
- the second matching degree may be characterized according to a second difference value between the feature of the candidate object and the second feature of the target object.
- determining the final matching degree between each object and the target object includes: determining each candidate according to the first difference value and the second difference value corresponding to the same candidate object The match difference between the object and the target object.
- filtering the object corresponding to the smallest final matching degree, and obtaining the target object includes: selecting the minimum matching difference value from the matching difference values corresponding to each candidate object; according to the minimum matching difference value, from Among the candidate objects in the current frame, objects that match the target object are filtered.
- the matching difference value is the difference value that exists when the candidate object matches the target object.
- the minimum matching difference value refers to the minimum matching difference value between the candidate object and the target object.
- the minimum matching difference value refers to the minimum matching difference value between the candidate object and the target object.
- the computer device may directly sum the first difference value and the second difference value corresponding to the same candidate object to obtain the matching difference value between each candidate object and the target object.
- the computer device may also perform non-summing other linear calculations or nonlinear calculations on the first difference value and the second difference value corresponding to the same candidate object to obtain the match between each candidate object and the target object. Difference value.
- the computer device can compare the multiple matching differences, and according to the comparison results, from Select the smallest matching difference value among multiple matching difference values.
- the candidate object corresponding to the minimum matching difference value has the smallest difference from the target object, and is also the closest. Therefore, the computer device can determine the candidate object corresponding to the minimum matching difference value in the current frame as a match with the target object. Therefore, the target object is tracked in the current frame, which is the candidate object corresponding to the smallest matching difference value in the current frame.
- the computer device may also compare the minimum matching difference value with a preset difference threshold. When the minimum matching difference value is less than or equal to the preset difference threshold, it is determined that the minimum matching difference value corresponds to the current frame
- the candidate for is the target. It can be understood that since there may be no target object in the current frame, the candidate object corresponding to the minimum matching difference value may not necessarily be the target object. Comparing the minimum matching difference value with the preset difference threshold can prevent Special circumstances occur, which further improves the accuracy of object tracking.
- Fig. 3 is a schematic diagram of the principle of an object tracking method in an embodiment.
- the first 6 frames of the current frame is equivalent to the historical frame.
- the frames before the current frame in the multi-frame image are called "historical frames". ", That is, it is the previous frame of the current frame, and now the target object G2 is to be tracked from the current frame.
- Computer equipment from the previous frame The feature of the target object G2 is extracted from, and the first feature of the target object G2 is obtained.
- from the first 6 historical frames as well as Extract the features of G2 from, and get 6 second features of the target object G2.
- the computer equipment needs to detect two candidate objects in the current frame (ie, object g1 (men) and object g2 (women)), and extract the features of these two candidates respectively to obtain the features of candidate object g1 and candidate object g2. feature. Then, the computer device can determine the first difference between the feature of the candidate object g1 and the first feature, and respectively determine the second difference between the feature of the candidate object g1 and the second feature, and then determine the first difference between the feature of the candidate object g1 and the second feature. The difference value determines the matching difference value h1 between the candidate object g1 and the target object G2 in the current frame.
- the computer device obtains the matching difference value h2 between the candidate object g2 and the target object G2 in the current frame according to the same method. If h2 is smaller than h1 and is within the preset difference threshold range, it can be determined that the candidate object g2 in the current frame is the target object G2 to be tracked.
- determining the matching difference value between each candidate object and the target object according to the first difference value and the second difference value corresponding to the same candidate object includes: obtaining a first weight corresponding to the first feature of the target object ; Obtain the second weight corresponding to each second feature; determine the first product of the first difference value corresponding to each candidate object and the first weight, and determine each second difference value corresponding to the candidate object and the corresponding second weight The second product corresponding to the same second feature corresponds to the second weight; the first product corresponding to the same candidate object and each second product are summed to obtain the difference between each candidate object and the target object The match difference value of.
- the first weight is used to indicate the degree of influence of the first feature of the target object on the object matching result.
- the second weight is used to indicate the degree of influence each second feature has on the object matching result.
- the first weight and the second weight may be preset fixed values.
- the first weight and the second weight can also be dynamically generated and determined according to actual conditions.
- each second feature may have the same preset second weight.
- Each second feature can also correspond to a different second weight.
- the computer device can obtain the first product of the first difference value corresponding to each candidate object and the first weight, and calculate the second product of each second difference value corresponding to the candidate object and the corresponding second weight.
- the computer device may sum the first product and each second product corresponding to the same candidate object to obtain the matching difference value between each candidate object and the target object.
- the degree of influence of the first feature and each second feature on the difference matching is taken into consideration, so that the determined matching difference value between each candidate object and the target object is more accurate, thereby improving the accuracy of object tracking.
- the current frame is the current frame of the current round; the first feature of the target object is extracted in the current round.
- Obtaining the first weight corresponding to the first feature of the target object includes: when obtaining the previous frame as the current frame of the previous round, the matching difference value corresponding to the object in the previous frame that matches the target object; The difference value determines the first weight corresponding to the first feature of the target object extracted in the current round.
- the object that matches the target object and the corresponding matching difference value (ie, the smallest matching difference value) of the matched object must also be calculated. Therefore, it can be based on the previous
- the matching difference value corresponding to the matched object calculated in one frame determines the first weight corresponding to the first feature of the target object involved in the calculation of the new current frame.
- the first weight is the matching confidence of the first feature of the target object.
- the matching confidence of the first feature of the target object refers to the matching difference value corresponding to the object in the previous frame that matches the target object when the previous frame is used as the current frame of the previous round.
- the computer device may directly use the matching difference value corresponding to the matched object calculated for the previous frame as the matching confidence corresponding to the first feature of the target object involved in the calculation of the new current frame, that is, to obtain The first weight.
- the computer device may also perform linear or non-linear transformation on the matching difference value corresponding to the matched object calculated for the previous frame to obtain the first target object involved in the calculation of the new current frame.
- the first weight corresponding to a feature may also perform linear or non-linear transformation on the matching difference value corresponding to the matched object calculated for the previous frame to obtain the first target object involved in the calculation of the new current frame.
- the current frame when the current frame is the second frame, since the previous frame is the first frame, and the matching object and the corresponding matching difference value are not calculated in the first frame, the current frame can be regarded as the first frame.
- the first weight corresponding to the first feature of the target object extracted in the previous frame may be the default value 1.
- the computer device may directly determine the second weight according to the first weight. For example, a computer device can subtract the first weight from 1 to obtain the second weight. In other embodiments, the computer device may also perform linear or nonlinear transformation on the first weight to obtain the second weight.
- the match difference value corresponding to the object in the previous frame that matches the target object is used to determine the target object extracted in the current round
- the first weight corresponding to the first feature of enables the first weight to be dynamically determined according to the historical tracking of the target object in the previous frame, which can more accurately determine the degree of influence of the first feature of the target object on the difference matching. To improve the accuracy of object tracking.
- obtaining the second weight corresponding to each second feature includes: obtaining the initial weight corresponding to each second feature; determining the weight coefficient according to the first weight; according to the initial weight and weight corresponding to each second feature, respectively The product of the coefficients obtains the second weight corresponding to each second feature.
- the initial weight is the initial weight.
- the initial weight corresponding to each second feature may be a preset value or dynamically determined according to actual conditions.
- the initial weights corresponding to the second features may be the same or different.
- the weight coefficient is determined according to the first weight to adjust the initial weight to obtain the second weight, which can improve the accuracy of the second weight, thereby improving the accuracy of object tracking.
- the current frame is the current frame of the current round; each second feature is extracted in the current round.
- obtaining the respective initial weights corresponding to the second features includes: obtaining the matching difference value corresponding to the object in each frame that matches the target object when each frame before the current frame is used as the current frame; For each obtained matching difference value, the initial weight corresponding to each second feature is determined.
- the initial weight corresponding to the second feature is obtained by dividing the second matching confidence of the second feature of the target object by the sum of all matching confidences.
- the weight coefficient is equal to the difference of 1 minus the first weight.
- the second matching confidence of the second feature of the target object refers to the matching difference value of the object matching the target object when the frame before the current frame corresponding to the second feature of the target object is used as the current frame.
- the sum of all matching confidences refers to the sum of the second matching confidences of the second features obtained by the target object in the frames before the current frame.
- the initial weights corresponding to the second features of each target object may be determined according to the following formula:
- the initial weight of the second feature of the i-th target object in the n-th historical frame ie, the frame before the current frame
- the second matching confidence of the second feature of the i-th target object in the n-th historical frame That is, the second matching confidence of the second feature of the i-th target object in the k-th historical frame among the N historical frames; That is, the sum of the second matching confidences of the second features of the i-th target object in all the acquired historical frames;
- N historical frames for the i-th target object and h is the abbreviation of historical "historical”.
- the matching difference between the candidate object and the target object can be calculated by the following formula:
- X i represents the feature of the i-th target object
- Z j is the feature of the j-th candidate object in the current frame
- ham(X i , Z j ) represents the matching difference between the j-th candidate object and the i-th target object value
- Is the first feature of the i-th target object in the previous frame of the current frame Is the first difference between the first feature of the i-th target object and the feature of the j-th candidate object in the previous frame of the current frame, Is the matching confidence of the first feature of the i-th target object (that is, the first weight of the first feature of the i-th target object)
- the initial weight of each second feature is dynamically determined based on the historical tracking of the target object in the historical frame, so that the degree of influence of each second feature on the difference matching can be determined more accurately, so that Improve the accuracy of object tracking.
- Step1 We enter the second frame (that is, enter the second image frame), we consider the simplest scenario: suppose there are five target objects in the first frame, and the five target objects in the first frame are all in the second In the frame, and there are no other target objects in the second frame. Then, let us remember that the five objects in the second frame are Z 1 , Z 2 , Z 3 , Z 4 , and Z 5 (Z is the candidate object).
- Step2 Now calculate the distance between the candidate object in the first frame and the second frame (for example: Euclidean distance), assuming that the feature of our target object is a vector of 1*m. Then the distance (ie, the first difference) between the i-th target object in the first frame and the j-th candidate object in the second frame is:
- the distance between all the target objects in the first frame and all the candidate objects in the second frame forms a matrix, namely
- dis x1z1 is another way of writing dis(X 1 , Z 1 ), and the same applies to other elements in the matrix.
- Step3 For the second frame, the matching confidence (ie, the first weight) of the first feature of the target object extracted from the first frame c is the initialized value 1, and the value of c is substituted into formula (1)
- the value calculated in ⁇ 1 ie the initial weight of the second feature of the target object in the first historical frame (ie the first frame)) is:
- Step5 Substituting the result obtained above into formula (2), the matching difference between the j-th candidate object and the i-th target object in the second frame is obtained as:
- [1,1,1,1,1] is the first weight.
- Step6 Assuming that we have successfully tracked here, and the target objects in the first frame and the second frame just correspond, then, for the convenience of description, will track the third frame, the first target object extracted from the second frame
- the matching confidence of features is simplified as:
- Step7 We input the third frame (that is, the third image frame), and also assume that there are 5 candidate objects in the third frame, and there is a one-to-one correspondence between the 5 candidate objects and the 5 target objects in the second image.
- the 5 candidate objects in the third frame we mark the 5 candidate objects in the third frame as: M 1 , M 2 , M 3 , M 4 , M 5
- the Z here is only to indicate the object in the second frame.
- the "Z” in Step 8 is equivalent to the "X” in Step 2.
- "M” is equivalent to the "Z” in Step 2. It is represented by a different serial number here, which is only used for distinguishing and does not prevent the application of the above formulas (1) and (2). Therefore, dis(Z i , M j ) is equivalent to the first feature of the i-th target object in the second frame of the previous frame and the j-th candidate object in the third frame when the third frame is used as the current frame The first difference between the features. It can be understood that dis z1m1 is another way of writing dis(Z 1 , M 1 ), and the same applies to other elements in the matrix.
- Step9 The value of the matching confidence c of the first feature of the target object extracted from the second frame is substituted into formula (1), and the value of ⁇ is calculated as:
- Step10 Then, substituting the result obtained above into formula (2), the matching difference between the jth candidate object and the ith target object in the third frame can be calculated as:
- It is the matrix formed by the product of the first weight and the first difference obtained for each target object when simultaneously tracking 5 target objects. It is a vector formed by the product of the second weight and the second difference obtained for each target object.
- steps in the flowchart of FIG. 2 are displayed in sequence as indicated by the arrows, these steps are not necessarily performed in sequence in the order indicated by the arrows. Unless specifically stated in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least some of the steps in FIG. 2 may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution of these sub-steps or stages The sequence is not necessarily performed sequentially, but may be performed alternately or alternately with other steps or at least part of the sub-steps or stages of other steps.
- an object tracking device 400 including: a feature extraction module 402, a feature matching module 404, and an object recognition module 406, wherein:
- the feature extraction module 402 is used to select the current frame from a multi-frame image; perform feature extraction on the target object in the previous frame of the current frame to obtain the first feature of the target object; from the multi-frame image located before the current frame In each frame, the feature of the target object is extracted respectively to obtain the second feature of the target object; the feature of each candidate object included in the current frame is extracted.
- the feature matching module 404 is used to match the feature of each candidate object with the first feature to obtain a first matching result, and to match the feature of each candidate object with the second feature to obtain a second matching result; according to the same candidate object
- the corresponding first matching result and second matching result determine the final matching result between each candidate object and the target object.
- the object recognition module 406 is used to identify the target object from the candidate objects in the current frame according to the final matching result.
- the current frame includes multiple objects
- the apparatus 400 further includes:
- the candidate object determination module 403 is configured to obtain at least one object detection template set in advance; respectively match each object included in the current frame with the object detection template; and determine the object successfully matched as a candidate object.
- the target object belongs to a pedestrian object;
- the object detection template includes a human body frame template.
- the candidate object determination module 403 is further configured to match each object included in the current frame with the human body frame template; according to the matching result, identify pedestrian objects included in the current frame as candidate objects.
- the candidate object determination module 403 is further configured to perform edge recognition on the object for each object included in the current frame to obtain the external contour feature of the object; match the external contour feature of the object with the human body frame template; When the matching is successful, the object is determined to be a pedestrian object.
- the target object belongs to a pedestrian object; the current frame includes multiple objects.
- the candidate object determination module 403 is also used for intercepting the object area map including the object from the current frame for each object in the current frame; the object occupies the main area of the object area map; and each object area map is input into the pre-trained human body.
- the recognition results for the objects included in each object area map are output; when the recognition results characterize the object as a human body, the object is determined as a candidate object.
- the first matching result includes a first matching degree between the object and the target object; the second matching result includes a second matching degree between the object and the target object.
- the feature matching module 404 is further configured to determine the final matching degree between each object and the target object according to the first matching degree and the second matching degree corresponding to the same object.
- the object recognition module 406 is also used to select the smallest final matching degree from the final matching degree corresponding to each object; from the objects in the current frame, filter the object corresponding to the smallest final matching degree to obtain the target object.
- Each module in the aforementioned object tracking device can be implemented in whole or in part by software, hardware, and a combination thereof.
- the foregoing modules may be embedded in the form of hardware or independent of the processor in the computer device, or may be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the foregoing modules.
- a computer device is provided.
- the computer device may be the server 120 in FIG. 1, and its internal structure diagram may be as shown in FIG. 6.
- the computer equipment includes a processor, a memory, and a network interface connected through a system bus.
- the memory includes a non-volatile storage medium and internal memory.
- the non-volatile storage medium of the computer device can store an operating system and a computer-readable storage medium. When the computer-readable storage medium is executed, it can cause the processor to execute an object tracking method.
- the processor of the computer equipment is used to provide calculation and control capabilities, and supports the operation of the entire computer equipment.
- a computer-readable storage medium may be stored in the internal memory, and when the computer-readable storage medium is executed by the processor, the processor may execute an object tracking method.
- the network interface of the computer equipment is used for network communication.
- Fig. 7 is an internal block diagram of a computer device in another embodiment.
- the computer equipment may be a terminal, and the computer equipment includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus.
- the processor of the computer device is used to provide calculation and control capabilities.
- the memory of the computer device includes a non-volatile or volatile storage medium and internal memory.
- the non-volatile or volatile storage medium stores an operating system and a computer-readable storage medium.
- the internal memory provides an environment for the operation of the operating system in the non-volatile storage medium and the computer-readable storage medium.
- the network interface of the computer device is used to communicate with an external terminal through a network connection.
- the computer-readable storage medium implements an object tracking method when executed by the processor.
- the display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen
- the input device of the computer equipment can be a touch layer covered on the display screen, or it can be a button, a trackball or a touchpad set on the housing of the computer equipment , It can also be an external keyboard, touchpad, or mouse.
- FIGS. 6 and 7 are only block diagrams of part of the structure related to the solution of the present application, and do not constitute a limitation on the computer equipment to which the solution of the present application is applied.
- the computer device may include more or fewer components than shown in the figures, or combine certain components, or have a different component arrangement.
- a computer device including a memory and one or more processors, the memory stores a computer-readable storage medium, and when the computer-readable storage medium is executed by the processor, the processor executes the above object tracking method A step of.
- the steps of the object tracking method may be the steps in the object tracking method of the foregoing embodiments.
- a computer-readable storage medium is provided, and the computer-readable storage medium is stored thereon.
- the processor executes the steps of the above object tracking method.
- the steps of the object tracking method may be the steps in the object tracking method of each of the foregoing embodiments.
- the computer-readable storage medium may be non-volatile or volatile.
- Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
- Volatile memory may include random access memory (RAM) or external cache memory.
- RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Channel (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Image Analysis (AREA)
Abstract
一种对象追踪方法,涉及人工智能技术,特别涉及图像处理技术,包括:从多帧图像中选取当前帧;对当前帧的前一帧中的目标对象进行特征提取,得到目标对象的第一特征;从多帧图像中的位于当前帧之前的各帧中,分别提取目标对象的特征,得到目标对象的第二特征;提取当前帧中包括的各候选对象的特征;分别将各候选对象的特征与第一特征进行匹配,得到第一匹配结果,并分别将各候选对象的特征与第二特征进行匹配,得到第二匹配结果;根据同一候选对象对应的第一匹配结果和第二匹配结果,确定各候选对象与目标对象之间的最终匹配结果;及按照最终匹配结果,从当前帧中的各候选对象中,识别目标对象。
Description
相关申请的交叉引用
本申请要求于2019年08月20日提交中国专利局,申请号为2019107695216,申请名称为“对象追踪方法、装置、计算机设备和存储介质”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
本申请涉及人工智能技术领域,尤其涉及图像处理技术领域,特别是涉及一种对象追踪方法、装置、计算机设备和存储介质。
随着科学技术的飞速发展,各种技术层出不穷。通过图像进行对象追踪的技术,在很多领域都起到很重要的作用。比如,警察通常需要通过监控视频来追踪犯罪嫌疑人,这种情况下,对象追踪技术就至关重要。
然而,发明人意识到,传统方法中,是将当前帧与前一帧进行匹配,从而根据匹配差异,从当前帧中确定出所要追踪的目标对象。这样一来,如果前一帧中的目标对象出现模糊或者部分被遮挡的情况,就会导致前一帧中的有效特征信息比较少,导致匹配不准确,进而导致对象追踪不准确。
发明内容
根据本申请公开的各种实施例,提供一种对象追踪方法、装置、计算机设备和存储介质。
一种对象追踪方法,包括:
从多帧图像中选取当前帧;
对当前帧的前一帧中的目标对象进行特征提取,得到目标对象的第一特征;
从所述多帧图像中的位于所述当前帧之前的各帧中,分别提取所述目标对象的特征,得到所述目标对象的第二特征;
提取所述当前帧中包括的各候选对象的特征;
将各候选对象的特征与所述第一特征进行匹配,得到第一匹配结果,并将各候选对象的特征与所述第二特征进行匹配,得到第二匹配结果;
根据同一候选对象对应的所述第一匹配结果和第二匹配结果,确定各候选对象与所述目标对象之间的最终匹配结果;及
按照最终匹配结果,从所述当前帧中的各候选对象中,识别目标对象。
一种对象追踪装置,包括:
特征提取模块,用于从多帧图像中选取当前帧;对当前帧的前一帧中的目标对象进行特征提取,得到目标对象的第一特征;从所述多帧图像中的位于所述当前帧之前的各帧中,分别提取所述目标对象的特征,得到所述目标对象的第二特征;提取所述当前帧中包括的各候选对象的特征;
特征匹配模块,用于将各候选对象的特征与所述第一特征进行匹配,得到第一匹配结果,并将各候选对象的特征与所述第二特征进行匹配,得到第二匹配结果;根据同一候选对象对应的所述第一匹配结果和第二匹配结果,确定各候选对象与所述目标对象之间的最 终匹配结果;及
对象识别模块,用于按照最终匹配结果,从所述当前帧中的各候选对象中,识别目标对象。
一种计算机设备,包括存储器和一个或多个处理器,所述存储器存储有计算机可读存储介质,计算机可读指令被处理器执行时,使得一个或多个处理器执行本申请各实施例中对象追踪方法中的步骤。
一个或多个存储有计算机可读指令的计算机可读存储介质,计算机可读指令被一个或多个处理器执行时,使得一个或多个处理器执行本申请各实施例中对象追踪方法中的步骤。上述对象追踪方法、装置、计算机设备和存储介质,分别获取前一帧中目标对象的第一特征,以及在当前帧之前的各帧中目标对象的第二特征,相当于,既考虑目标对象在较近的前一帧中的第一特征,又考虑到目标对象在之前的多帧中的第二特征,从而使得所提取的目标特征的信息量更多,更加的准确。进而,将当前帧中各候选对象的特征分别与目标对象的第一特征和各第二特征进行匹配,根据第一匹配结果和第二匹配结果确定候选对象与目标对象之间的最终匹配结果,从而能够使各候选对象和目标对象之间的匹配结果更加准确,基于该匹配结果能够从当前帧中更加准确地识别出目标对象,进而实现了对目标对象更加准确地追踪。
本申请的一个或多个实施例的细节在下面的附图和描述中提出。本申请的其它特征和优点将从说明书、附图以及权利要求书变得明显。
图1为根据一个或多个实施例中对象追踪方法的应用场景图;
图2为根据一个或多个实施例中对象追踪方法的流程示意图;
图3为根据一个或多个实施例中对象追踪方法的原理示意图;
图4为根据一个或多个实施例中对象追踪装置的框图;
图5为根据一个或多个实施例中追踪匹配模块的框图;
图6为根据一个或多个实施例中计算机设备的框图;
图7为另一个实施例中计算机设备的框图。
本申请提供的对象追踪方法,可以应用于如图1所示的应用环境中。终端110通过网络与服务器120通过网络进行通信。终端110可以但不限于是各种个人计算机、笔记本电脑、智能手机、平板电脑和便携式可穿戴设备,服务器120可以用独立的服务器或者是多个服务器组成的服务器集群来实现。可以理解,终端110具备图像采集功能。
终端110可以采集视频,并将视频发送至服务器120,由服务器120根据视频执行本申请各实施例中的对象追踪方法。服务器120可以从视频的多帧图像中选取当前帧,对当前帧的前一帧中的目标对象进行特征提取,得到目标对象的第一特征;从所述多帧图像中的位于所述当前帧之前的各帧中,分别提取所述目标对象的特征,得到所述目标对象的第二特征;提取所述当前帧中包括的各候选对象的特征;将各候选对象的特征与所述第一特征进行匹配,得到第一匹配结果,并将各候选对象的特征与所述第二特征进行匹配,得到第二匹配结果;根据同一候选对象对应的所述第一匹配结果和第二匹配结果,确定各候选对象与所述目标对象之间的最终匹配结果;按照最终匹配结果,从所述当前帧中的各候选 对象中,识别目标对象。进一步地,服务器120可以在当前帧中将识别出的目标对象添加突出显示标记,并通知终端110,以指示终端110在当前帧中将目标对象进行突出显示。
需要说明的是,在其他实施例中,终端110自身也可以对各帧图像执行本申请各实施例中的对象追踪方法,而不需要发送至服务器120。这里对谁执行对象追踪方法不做限定。
在一些实施例中,如图2所示,提供了一种对象追踪方法,以该方法应用于计算机设备为例进行说明,计算机设备可以是图1中的服务器120,包括以下步骤:
S202,从多帧图像中选取当前帧。
当前帧,是当前处理的图像帧。
在一些实施例中,多帧图像可以是视频(比如,监控视频)里面的多个图像。
需要说明的是,该对象追踪方法也可以不限定于视频中追踪对象的场景,可以适用于在任何多帧图像中追踪对象的场景。所以,在其他实施例中,多帧图像,也可以是多张图片。
S204,对当前帧的前一帧中的目标对象进行特征提取,得到目标对象的第一特征。
当前帧的前一帧,是多帧图像中的相邻于当前帧的前面一帧图像帧,即,指多帧图像中的在当前帧之前的最近一帧。目标对象,是指需要进行追踪的对象。
可以理解,目标对象存在于前一帧中,需要在当前帧中识别出该目标对象,从而实现对该目标对象从前一帧到当前帧的追踪。
目标对象的第一特征,即为目标对象在当前帧的前一帧中的特征。
需要说明的是,目标对象可以为一个或多个。当目标对象为多个时,则可以针对每个目标对象分别执行步骤S204~S206、以及S210~S214的步骤,以实现在当前帧中追踪每个目标对象。
具体地,计算机设备可以从当前帧的前一帧中,识别出目标对象,并通过卷积处理或者边缘、骨骼架构以及色彩检测等至少一种检测处理,来提取目标对象的特征,得到目标对象的第一特征。可以理解,由于目标对象的第一特征是从最接近于该当前帧的前一帧中提取到的,所以,第一特征相当于目标对象的近期特征。
S206,从多帧图像中的位于当前帧之前的各帧中,分别提取目标对象的特征,得到目标对象的第二特征。
多帧图像中的位于当前帧之前的各帧,是指多帧图像中的位于当前帧之前的图像帧。可以理解,当多帧图像是视频帧中的多个图像帧时,位于当前帧之前的各帧,即为多帧图像中的在当前帧之前产生的视频图像帧。
具体地,计算机设备可以从多帧图像中的位于当前帧之前的各帧中,分别检测出目标对象,并从中分别提取目标对象的特征,得到与各帧分别对应的目标对象的第二特征。同样地,计算机设备可以针对位于当前帧之前的各帧,通过卷积处理或者边缘、骨骼架构以及色彩检测等检测处理,来提取目标对象的特征,得到目标对象的各第二特征。
需要说明的是,当多帧图像中的位于当前帧之前的帧中不包括目标对象时,则无法从中提取出目标对象的第二特征,则在后续步骤S210的处理中,使用不到该帧。因此,本申请各实施例中,是基于从多帧图像中的位于当前帧之前的各帧中提取出的目标对象的第二特征进行处理的,未提取到目标对象的第二特征的情况自然不在考虑范围内。
可以理解,目标对象的第二特征是从多帧图像中的位于所述当前帧之前的各帧中提取 到的,由于在当前帧之前的各帧已经执行过对象追踪方法,所以,其相对于当前帧来说属于历史帧(即,已经执行过本申请各实施例中的对象追踪方法的图像帧),所以,从位于所述当前帧之前的各帧中提取的目标对象的第二特征,相当于目标对象的历史特征。
可以理解,可以是从多帧图像中的位于当前帧之前的全部或部分(即至少部分)图像帧中,分别提取目标对象的特征,得到目标对象的第二特征。
在一些实施例中,计算机设备可以选取多帧图像中的、且位于当前帧的前预设数量的帧,并从选取的各帧中,分别提取目标对象的特征,得到目标对象的第二特征。可以理解,是对选取的每帧都提取目标对象的特征,因而能够得到与选取的每帧分别对应的目标对象的第二特征。
具体地,计算机设备也可以从当前帧起按序往前选取预设数量的帧。比如,当前帧为视频中的第5张图像帧,预设数量为3,则选取的历史帧则为从第5张图像帧起往前选3张图像帧,即,第2~4张图像帧为选取的图像帧,然后从选取的第2~4张图像帧中,分别提取目标对象的特征,得到与第2~4张图像帧分别对应的目标对象的第二特征。那么,该实施例中,选取的帧中则包括当前帧的前一帧。可以理解,当预设数量大于1时,前预设数量的历史帧中除包括当前帧的前一帧,还包括在前一帧之前的帧。
在其他实施例中,计算机设备可以在位于当前帧的各帧中,随机选取满足预设数量的帧。比如,当前帧为视频中的第5张图像帧,预设数量为3,则可以随机选第1张、第3张和第4张图像帧,然后从选取的第1、3和4张图像帧中,分别提取目标对象的特征,得到与第1、3和4张图像帧分别对应的目标对象的第二特征。
S208,提取当前帧中各候选对象的特征。
候选对象,是指当前帧中能够观测到的用于被判断是否为目标对象的对象。候选对象为至少一个。
在一些实施例中,候选对象可以包括人、车辆、动物和物品等至少一种。
在一些实施例中,计算机设备可以将当前帧中所有的对象都识别为候选对象。比如,当前帧中有两个人、一条狗以及一辆车,计算机设备可以不考虑对象的类别,直接将当前帧中的对象都作为候选对象,即,将两个人、一条狗以及一辆车都作为候选对象。
在另一个实施例中,计算机设备也可以从当前帧中能够观测到的对象中,选取部分对象作为候选对象。具体地,计算机设备可以获取目标对象的类别,从当前帧中能够观测到的对象中,选取与目标对象的类别相符的对象作为候选对象。比如,当前帧中有两个人、一条狗以及一辆车,假设目标对象的类别为“人”,那么则可以选取当前帧中的两个“人”作为候选对象。
具体地,计算机设备可以对每个候选对象进行边缘检测(外形维度)、骨骼架构检测(内部架构维度)和颜色检测(色彩维度)等至少一种维度的检测,从而得到各候选对象的特征。
在一些实施例中,计算机设备可以对每个候选对象进行卷积处理,通过多轮卷积处理,提取候选对象的特征。
S210,将各候选对象的特征与所述第一特征进行匹配,得到第一匹配结果,并将各候选对象的特征与所述第二特征进行匹配,得到第二匹配结果。
具体地,计算机设备可以将每个候选对象的特征与目标对象的第一特征进行匹配,得 到与每个候选对象对应的第一匹配结果。计算机设备可以将每个候选对象的特征与目标对象的第二特征进行匹配,得到与每个候选对象对应的第二匹配结果。
在一些实施例中,第一匹配结果可以包括匹配成功和匹配失败。第二匹配结果也可以包括匹配成功和匹配失败。即,匹配结果可以不为数值形式,而是直接地结论描述形式。
在另一个实施例中,第一匹配结果,包括候选对象的特征与目标对象的第一特征之间的第一匹配度。第二匹配结果,包括候选对象的特征与目标对象的第二特征之间的第二匹配度。
在一些实施例中,第一匹配度,可以根据候选对象的特征与目标对象的第一特征之间的第一差异值进行表征。即,第一差异值,用于表征各候选对象的特征与目标对象的第一特征之间的特征匹配度。第二匹配度,可以根据候选对象的特征与目标对象的第二特征之间的第二差异值进行表征。即,第二差异值,用于表征各候选对象的特征分别与目标对象的各第二特征之间的特征匹配度。
可以理解,由于位于所述当前帧之前的帧为至少一个。所以,当位于所述当前帧之前的帧为多个时,从多个帧中可以分别提取相对应的第二特征,即有多个目标对象的第二特征。那么,各候选对象的特征分别与目标对象的每个第二特征之间都具有第二差异值,即可以得到多个第二差异值。
在一些实施例中,针对每个候选对象,计算机设备可以求取该候选对象的特征与目标对象的第一特征之间的欧式距离,从而得到第一差异值,以及分别求取该候选对象的特征与目标对象的各第二特征之间的欧式距离,得到该候选对象的特征与目标对象的各第二特征之间的第二差异值。
S212,根据同一候选对象对应的所述第一匹配结果和第二匹配结果,确定各候选对象与所述目标对象之间的最终匹配结果。
在一些实施例中,当同一候选对象对应的第一匹配结果和第二匹配结果皆为匹配成功时,则确定最终匹配结果为该候选对象为目标对象。
在另一个实施例中,计算机设备也可以根据对应于同一候选对象的第一匹配度和第二匹配度,确定各候选对象对应的最终特征匹配度。该最终特征匹配度,即为最终匹配结果。可以理解,当第一匹配度和第二匹配度分别由第一差异值和第二差异值进行表征时,计算机设备也可以根据对应于同一候选对象的第一差异值和第二差异值,确定各候选对象与目标对象之间的匹配差异值。可以理解,各候选对象与目标对象之间的匹配差异值,即为候选对象与所述目标对象之间的最终匹配结果。
S214,按照最终匹配结果,从所述当前帧中的各候选对象中,识别目标对象。
在一些实施例中,针对每个候选对象,当该候选对象对应的最终匹配结果为该候选对象为目标对象时,则从当前帧中识别出该候选对象,得到目标对象。
在另一个实施例中,计算机设备可以根据各候选对象与与目标对象之间的匹配差异值,从当前帧中的各候选对象中,识别目标对象。比如,从当前帧中的各候选对象中,选取匹配差异值最小的候选对象,得到目标对象。
在一些实施例中,计算机设备还可以将筛选出的对象在当前帧中进行突出显示,以体现该对象为所追踪的目标对象。
上述对象追踪方法,分别获取前一帧中目标对象的第一特征,以及在当前帧之前的各 帧中目标对象的第二特征,相当于,既考虑目标对象的近期特征,又考虑到目标对象的历史特征,从而使得所提取的目标特征的信息量更多,更加的准确。进而,将当前帧中各候选对象的特征分别与目标对象的第一特征和各第二特征进行匹配,根据第一匹配结果和第二匹配结果确定候选对象与目标对象之间的最终匹配结果,从而能够使各候选对象和目标对象之间的匹配结果更加准确,基于该匹配结果能够从当前帧中更加准确地识别出目标对象,进而实现了对目标对象更加准确地追踪。
在一些实施例中,步骤S208提取当前帧中各候选对象的特征包括:获取预先设置的至少一个对象检测模板;分别将当前帧中包括的各对象与对象检测模板进行匹配;确定匹配成功的对象为候选对象。
对象检测模板,是预先设置的用于检测对象的模板。
具体地,计算机设备可以预先设置对象检测模板。计算机设备可以将当前帧中的对象与对象检测模板进行匹配,将当前帧中与该对象检测模板匹配的对象作为候选对象。进而,从当前帧中提取候选对象的特征。
在一些实施例中,可以以对象的类别为维度,设置相应的对象检测模板。即,一种对象检测模板用于检测一种类别的对象。
比如,要检测当前帧中的人,那么对象检测模板可以为人体框架模板,计算机设备可以检测当前帧中对象的框架,将该框架与人体框架模板进行匹配,如果匹配成功,则说明该对象为人,从而将当前帧中所有的人都检测识别出来。那么,所识别出的所有的人,则为候选对象。
又比如,要检测当前帧中的车辆,那么对象检测模板可以为车辆框架模板(即能判别出是车辆的特征的模板),计算机设备可以将当前帧中对象与该模板进行匹配,匹配成功则说明该对象为车辆,从而识别出当前帧中的所有车辆。那么,所识别出的所有的车辆,就为候选对象。
在一些实施例中,还可以根据二级分类,设置相应的对象检测模板。假设,针对“人”这一类别,可以进行二级分类为“老人”、“小孩”、“男人”和“女人”等,那么,就可以对应二级分类设置对象检测模板,比如,设置小孩检测模板,那么,就可以从当前帧中识别出所有的小孩。这样一来,就可以用来自动寻找走失的儿童。
在一些实施例中,计算机设备可以获取目标对象的类别,获取与目标对象的类别对应的对象检测模板,然后将当前帧中的对象与对象检测模板进行匹配,将当前帧中与该对象检测模板匹配的对象作为候选对象。
假设,目标对象的类别为“人”,那么,就获取用于检测出“人”的对象检测模板,比如,人体框架模板,从而检测出当前帧中的人,作为候选对象。
上述实施例中,通过对象检测模板可以对当前帧中的对象进行筛选,以筛选出符合要求的对象作为候选对象,这样一来,就不用针对当前帧中的所有对象都与目标对象进行差异匹配,从而在保证对象追踪正确性的同时,节约了计算资源。
在一些实施例中,所述目标对象属于行人对象;所述对象检测模板包括人体框架模板。本实施例中,分别将当前帧中包括的各对象与所述对象检测模板进行匹配包括:分别将当前帧中包括的各对象与所述人体框架模板进行匹配。确定匹配成功的对象为候选对象包括:根据匹配结果,识别所述当前帧中包括的行人对象,作为候选对象。
行人对象,即为图像帧中展示的人形图像内容。人体框架模板,是预先设置的人体框架。可以理解,人体框架,是包含了人体存在共性的架构特征。即,除特殊人群之外,正常的人通常情况下能够与该人体框架匹配。
具体地,计算机设备可以先检测出当前帧中包括的对象,然后将各对象分别与所述人体框架模板进行匹配,根据匹配结果,识别所述当前帧中包括的行人对象,作为候选对象。可以理解,当匹配结果为与该对象人体框架模板匹配,则识别该对象为行人对象,那么,该对象即为候选对象。
可以理解,人体框架模板可以从轮廓角度进行设置,即,人体框架模板包括人体共性轮廓,计算机设备可以将当前帧中的各对象的外形轮廓与人体框架模板进行匹配。
在一些实施例中,所述分别将当前帧中包括的各对象与所述人体框架模板进行匹配包括:针对当前帧中包括的每个对象,对所述对象进行边缘识别,得到所述对象的外部轮廓特征;将所述对象的外部轮廓特征与所述人体框架模板进行匹配;当匹配成功时,则判定所述对象为行人对象。
外部轮廓特征,是对象的外部轮廓的特征,用于表征对象的外部轮廓。
具体地,计算机设备可以通过边缘检测处理,检测当前帧中包括的各对象的边缘特征点,将边缘特征点按序连接,得到外部轮廓线,即为外部轮廓特征数据。计算机设备可以将得到的各对象的外部轮廓特征数据与包括人体共性轮廓的人体框架模板进行匹配。当匹配成功时,则判定该对象为行人对象。
在其他实施例中,人体框架模板还可以从人体骨骼角度进行设置,即人体框架模板为人体骨骼框架模板。可以理解,人身体上的骨头差异不大,所以具有一定的人体共性,因此可以设置人体骨骼框架模板。人体骨骼框架模板中包括预设的外观骨骼关键点。外观骨骼关键点,是对外展现的、能够从外观直接看到的骨骼点。那么,计算机设备可以将当前帧中各对象与包括预设的外观骨骼关键点的人体骨骼框架模板进行匹配,当匹配时,则判定该对象为行人对象,即为候选对象。
在一些实施例中,目标对象属于行人对象;当前帧中包括多个对象。在步骤S208之前,该方法还包括:针对当前帧中的每个对象,从当前帧中截取包括对象的对象区域图;对象占据对象区域图中的主要区域;分别将各个对象区域图输入预先训练的人体识别模型中,输出针对各对象区域图中包括的对象的识别结果;当识别结果表征对象为人体时,则判定对象为候选对象。
对象区域图,是对象占主要区域的图像。可以理解,对象区域图,是从当前帧中截取的一部分。比如,当前帧中包括3个对象A~C,从中截取A的对象区域图,则A的对象区域图中,A占主要区域。
人体识别模型,是预先训练的用于识别人的机器学习模型。人体识别模型,可以预先通过样本数据进行迭代地机器学习训练得到。样本数据包括样本图像和对象类别标记。正样本图像中的图像内容包括人,负样本图像中的图像内容包括除人以外的对象。正样本图像对应的对象类别标记为人体标记,正样本图像对应的对象类别标记为非人体标记。
计算机设备可以分别将各个对象区域图输入预先训练的人体识别模型中,输出各对象区域图对应的对象的识别结果;当识别结果表征该对象为人体时,则判定对象为候选对象。当识别结果表征该对象为非人体时,则判定该对象不为候选对象。
在一些实施例中,第一匹配结果包括对象与目标对象之间的第一匹配度;第二匹配结果包括对象与目标对象之间的第二匹配度。本实施例中,步骤S212包括:根据对应于同一对象的第一匹配度和第二匹配度,确定各对象与目标对象之间的最终匹配度。步骤S214包括:从各对象所对应的最终匹配度中选取最小的最终匹配度;从当前帧的各对象中,筛选最小的最终匹配度所对应的对象,得到目标对象。
在一些实施例中,第一匹配度,可以根据候选对象的特征与目标对象的第一特征之间的第一差异值进行表征。第二匹配度,可以根据候选对象的特征与目标对象的第二特征之间的第二差异值进行表征。根据对应于同一对象的第一匹配度和第二匹配度,确定各对象与目标对象之间的最终匹配度包括:根据对应于同一候选对象的第一差异值和第二差异值,确定各候选对象与目标对象之间的匹配差异值。从当前帧的各对象中,筛选最小的最终匹配度所对应的对象,得到目标对象包括:从各候选对象所对应的匹配差异值中选取最小匹配差异值;根据所述最小匹配差异值,从当前帧的各候选对象中,筛选与所述目标对象匹配的对象。
匹配差异值,是候选对象与目标对象进行匹配时所存在的差异值。最小匹配差异值,是指候选对象与目标对象之间的匹配差异值最小。
可以理解,匹配差异值越小,候选对象与目标对象差异越小,越接近;反之,匹配差异值越大,候选对象与目标对象差异越大,越不相同。
最小匹配差异值,是指候选对象与目标对象之间的匹配差异值最小。
在一些实施例中,计算机设备可以直接将对应于同一候选对象的第一差异值和第二差异值求和,得到各候选对象与目标对象之间的匹配差异值。
在其他实施例中,计算机设备也可将对应于同一候选对象的第一差异值和第二差异值做非求和的其他线性计算或非线性计算,得到各候选对象与目标对象之间的匹配差异值。
可以理解,当候选对象为多个(即大于或等于两个)时,得到的匹配差异值也为多个,计算机设备可以将多个匹配差异值进行大小比对,并根据比对结果,从多个匹配差异值中选取最小匹配差异值。最小匹配差异值所对应的候选对象与目标对象差异最小,也就最接近。因此,计算机设备可以将最小匹配差异值在当前帧中所对应的候选对象,判定为与目标对象匹配。从而在当前帧中追踪到该目标对象,即为最小匹配差异值在当前帧中所对应的候选对象。
在其他实施例中,计算机设备也可以将最小匹配差异值与预设差异阈值进行比对,当最小匹配差异值小于或等于预设差异阈值,则判定该最小匹配差异值在当前帧中所对应的候选对象为目标对象。可以理解,由于当前帧中也可以存在不存在目标对象的情况,所以,最小匹配差异值所对应的候选对象也不一定是目标对象,将最小匹配差异值与预设差异阈值比对,能够防止特殊情况发生,进一步地提高对象追踪的准确性。
图3为一个实施例中对象追踪方法的原理示意图。参照图3,
以及
即为当前帧的前6个帧,可以理解,当前帧之前的帧,相当于历史帧,以下为了表述简练,就将多帧图像中的位于所述当前帧之前的各帧称为“历史帧”,
即为当前帧的前一帧,现要从当前帧中追踪目标对象G2。计算机设备则要从前一帧
中提取该目标对象G2的特征, 得到目标对象G2的第一特征。并且,分别从前6个历史帧
以及
中提取G2的特征,得到目标对象G2的6个第二特征。计算机设备则要检测当前帧中的2个候选对象(即对象g1(男士)和对象g2(女士)),并分别提取这两个候选对象的特征,得到候选对象g1的特征和候选对象g2的特征。然后,计算机设备可以确定候选对象g1的特征与第一特征之间的第一差异值,以及分别确定候选对象g1的特征与第二特征的第二差异值,然后根据第一差异值和第二差异值,确定当前帧中候选对象g1和目标对象G2之间的匹配差异值h1。计算机设备按照同样地方法,求得当前帧中候选对象g2和目标对象G2之间的匹配差异值h2。h2小于h1且在预设差异阈值范围内,则可以判定当前帧中的候选对象g2即为所要追踪的目标对象G2。
在一些实施例中,根据对应于同一候选对象的第一差异值和第二差异值,确定各候选对象与目标对象之间的匹配差异值包括:获取目标对象的第一特征对应的第一权重;获取各第二特征对应的第二权重;确定各候选对象所对应的第一差异值和第一权重的第一乘积,以及确定候选对象所对应的各第二差异值与相对应第二权重的第二乘积;对应于同一第二特征的第二差异值和第二权重相对应;将对应于同一候选对象的第一乘积和各第二乘积求和,得到各候选对象与目标对象之间的匹配差异值。
第一权重,用于表示目标对象的第一特征对对象匹配结果所产生的影响程度。第二权重,用于表示各第二特征对对象匹配结果所产生的影响程度。
在一些实施例中,第一权重和第二权重,可以是预先设置的固定值。
在另一个实施例中,第一权重和第二权重,也可以根据实际情况动态生成决定。
可以理解,各第二特征可以具有预设的相同的第二权重。各第二特征也可以对应不同的第二权重。
计算机设备可以求取各候选对象所对应的第一差异值和第一权重的第一乘积,以及计算候选对象所对应的各第二差异值与相对应第二权重的第二乘积。计算机设备可以将对应于同一候选对象的第一乘积和各第二乘积求和,得到各候选对象与目标对象之间的匹配差异值。
上述实施例中,考虑到了第一特征和各第二特征对差异匹配的影响程度,因而使得确定出的各候选对象与目标对象之间的匹配差异值更加准确,从而提高对象追踪的准确性。
在一些实施例中,当前帧为当前轮次的当前帧;目标对象的第一特征,是在当前轮次中提取得到。获取目标对象的第一特征对应的第一权重包括:获取前一帧作为前一轮次的当前帧时,前一帧中的、且与目标对象匹配的对象所对应的匹配差异值;根据匹配差异值,确定与当前轮次中所提取的目标对象的第一特征对应的第一权重。
可以理解,前一帧作为当前帧时,同样要计算出其中与目标对象相匹配的对象以及该匹配的对象相应的匹配差异值(即,最小的匹配差异值),所以,就可以根据针对前一帧计算得到的所匹配的对象对应的匹配差异值,确定新的当前帧计算时所涉及到的目标对象的第一特征对应的第一权重。
在一些实施例中,第一权重,即为目标对象的第一特征的匹配置信度。目标对象的第一特征的匹配置信度,是指前一帧作为前一轮次的当前帧时,前一帧中的、且与目标对象匹配的对象所对应的匹配差异值。具体地,计算机设备可以直接将针对前一帧计算得到的所匹配的对象对应的匹配差异值,作为新的当前帧计算时所涉及到的目标对象的第一特征 对应的匹配置信度,即得到第一权重。
在另一个实施例中,计算机设备也可以将针对前一帧计算得到的所匹配的对象对应的匹配差异值进行线性或非线性变换,得到新的当前帧计算时所涉及到的目标对象的第一特征对应的第一权重。
在一些实施例中,当当前帧为第二帧时,由于其前一帧为第一帧,而第一帧并没有计算匹配的对象以及对应的匹配差异值,所以,可以当当前帧为第二帧时,其前一帧中提取的目标对象的第一特征对应的第一权重可以为默认值1。
为了便于理解,现举例说明。假设,第二帧为当前帧时,计算第二帧中3个候选对象与目标对象的匹配差异值(匹配差异值1、匹配差异值2以及匹配差异值3),从三个匹配差异值中选最小的匹配差异值3,将最小的匹配差异值3所对应的候选对象判定为与目标对象匹配。那么,接着,在当第三帧为当前帧时,则可以将前面计算得到的匹配差异值3作为目标对象的第一特征的第一权重。
在一些实施例中,计算机设备可以根据第一权重,直接确定第二权重。比如,计算机设备可以用1减去第一权重,得到第二权重。在其他实施例中,计算机设备也可以对第一权重进行线性或非线性变换,得到第二权重。
上述实施例中,根据前一帧作为前一轮次的当前帧时,前一帧中的、且与目标对象匹配的对象所对应的匹配差异值,确定与当前轮次中所提取的目标对象的第一特征对应的第一权重,使得第一权重能够根据对目标对象在前一帧中的历史追踪情况来动态确定,能够更加准确地确定目标对象的第一特征对差异匹配的影响程度,以提高对象追踪的准确性。
在一些实施例中,获取各第二特征对应的第二权重包括:获取各第二特征分别对应的初始权重;根据第一权重,确定权重系数;根据各第二特征分别对应的初始权重和权重系数的乘积,得到各第二特征对应的第二权重。
初始权重,即为初始化的权重。
在一些实施例中,各第二特征分别对应的初始权重可以为预设值或根据实际情况动态确定。各第二特征分别对应的初始权重,可以相同也可以不同。
上述实施例中,根据第一权重确定权重系数,以对初始权重进行调整得到第二权重,能够提高第二权重的准确性,从而提高对象追踪的准确性。
在一些实施例中,当前帧为当前轮次的当前帧;各第二特征,是在当前轮次中提取得到。本实施例中,获取各第二特征分别对应的初始权重包括:获取位于当前帧之前的各帧作为当前帧时,各帧中的、且与目标对象匹配的对象所对应的匹配差异值;根据获取的各匹配差异值,确定各第二特征分别对应的初始权重。
在一些实施例中,第二特征对应的初始权重,是通过该目标对象的第二特征的第二匹配置信度除以所有匹配置信度的总和得到。权重系数,等于1减去第一权重的差值。
目标对象的第二特征的第二匹配置信度,是指该目标对象的第二特征所对应的位于当前帧之前的帧作为当前帧时,与目标对象匹配的对象所对应的匹配差异值。所有匹配置信度的总和,是指目标对象在获取的位于当前帧之前的各帧中的第二特征的第二匹配置信度之和。
在一些实施例中,各目标对象的第二特征分别对应的初始权重可以根据以下公式确定:
即为第i个目标对象在第n个历史帧(即,位于当前帧之前的帧)中的第二特征的初始权重;
即为第i个目标对象在第n个历史帧中的第二特征的第二匹配置信度;
即为第i个目标对象在N个历史帧中的第k个历史帧中的第二特征的第二匹配置信度;
即为第i个目标对象在所获取的所有历史帧中的第二特征的第二匹配置信度之和;
为第i个目标对象有N个历史帧,h即为历史的“historical”的缩写。
在一些实施例中,可以通过以下公式来计算候选对象与目标对象之间的匹配差异值:
X
i代表第i个目标对象的特征,Z
j为当前帧中第j个候选对象的特征,ham(X
i,Z
j)表示第j个候选对象与第i个目标对象之间的匹配差异值;
为当前帧的前一帧中第i个目标对象的第一特征,
为当前帧的前一帧中第i个目标对象的第一特征和第j个候选对象的特征之间的第一差异,
为第i个目标对象的第一特征的匹配置信度(即为第i个目标对象的第一特征的第一权重);
代表第i个目标对象的第n个历史帧的第二特征,
当前帧的第n个历史帧中第i个目标对象的第二特征和第j个候选对象的特征之间的第二差异,
表示第i个目标对象在第n个历史帧中的第二特征的初始权重。
即为权重系数。可以理解,公式(2)中
与
的乘积,即相当于第i个目标对象在第n个历史帧中的第二特征的第二权重。
上述实施例中,各第二特征的初始权重,是根据对目标对象在历史帧中的的历史追踪情况来动态确定,进而能够更加准确地确定各第二特征对差异匹配的影响程度,从而能够提高对象追踪的准确性。
为了便于理解上述公式(2),下面通过具体例子来阐述(由于第一帧没有历史帧,即,第一帧没有之前的帧,所以没有讨论的必要,故省略第一帧,从第二帧开始阐述):
Step1:我们输入第二帧(即输入第二个图像帧),我们考虑最简单的场景:假设第一帧中有五个目标对象,而且,第一帧中的五个目标对象均在第二帧中,且第二帧中除此之外没有其他的目标对象。那么,我们记第二帧中的5个对象分别为Z
1,Z
2,Z
3,Z
4,Z
5(Z即为候选对象)。
Step2:现在计算第一帧和第二帧候选对象的距离(例如:欧式距离),假设我们目标对象的特征为1*m的一个向量。则第一帧的第i个目标对象和第二帧的第j个候选对象之间的距离(即第一差异)为:
则第一帧中所有的目标对象和第二帧中所有的候选对象之间距离(即第一差异)组成一个矩阵,即
可以理解,dis
x1z1即为dis(X
1,Z
1)的另一种写法,同理,也适用于矩阵中的其他元素。
Step3:对于第二帧来说,从第一帧提取的目标对象的第一特征的匹配置信度(即第一权重)c的值为初始化的值1,将该c的值代入公式(1)中计算ω
1(即目标对象在第1个历史帧(即第一帧)中的第二特征的初始权重)的值为:
Step4:则代入公式(2)中,得到:
Step5:将前面求得的结果,代入公式(2)中,求得第二帧中第j个候选对象与第i个目标对象之间的匹配差异值为:
其中,[1,1,1,1,1]为第一权重。
Step7:我们输入第三帧(即第三帧图像帧),同样假设第三帧中有5个候选对象,且5个候选对象和第二帧图像中的5个目标对象存在一一对应关系。为了记述方便,我们标记第三帧中的5个候选对象分别为:M
1,M
2,M
3,M
4,M
5
Step8:则此时公式(2)中:
需要说明的是,这里的Z只为了表示是第二帧中的对象,实际上,在对第三帧进行对象追踪处理时,Step8中的“Z”就相当于Step 2中的“X”,“M”相当于Step 2中的“Z”,这里用不同的序号表示,仅用于区分表示,并不妨碍对上述公式(1)和(2)的套用。因此,dis(Z
i,M
j)相当于第三帧作为当前帧时,作为其前一帧的第二帧中第i个目标对象的第一特征和该第三帧中第j个候选对象的特征之间的第一差异。可以理解,dis
z1m1即为dis(Z
1,M
1)的另一种写法,同理,也适用于矩阵中的其他元素。
Step9:将从第二帧中提取的目标对象的第一特征的匹配置信度c的值代入公式(1)中,计算ω的值为:
Step10:那么,将上面求得的结果,代入公式(2)中,可计算第三帧中第j个候选对象与第i个目标对象之间的匹配差异值为:
应该理解的是,虽然图2的流程图中的各个步骤按照箭头的指示依次显示,但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明,这些步骤的执行并没有严格的顺序限制,这些步骤可以以其它的顺序执行。而且,图2中的至少一部分步骤可以包括多个子步骤或者多个阶段,这些子步骤或者阶段并不必然是在同一时刻执行完成,而是可以在不同的时刻执行,这些子步骤或者阶段的执行顺序也不必然是依次进行,而是可与其它步骤或者其它步骤的子步骤或者阶段的至少一部分轮流或者交替执行。
在一些实施例中,如图4所示,提供了一种对象追踪装置400,包括:特征提取模块402、特征匹配模块404以及对象识别模块406,其中:
特征提取模块402,用于从多帧图像中选取当前帧;对当前帧的前一帧中的目标对象进行特征提取,得到目标对象的第一特征;从多帧图像中的位于当前帧之前的各帧中,分别提取目标对象的特征,得到目标对象的第二特征;提取当前帧包括的各候选对象的特征。
特征匹配模块404,用于将各候选对象的特征与第一特征进行匹配,得到第一匹配结果,并将各候选对象的特征与第二特征进行匹配,得到第二匹配结果;根据同一候选对象对应的第一匹配结果和第二匹配结果,确定各候选对象与目标对象之间的最终匹配结果。
对象识别模块406,用于按最终匹配结果,从当前帧的各候选对象中,识别目标对象。
如图5所示,在一些实施例中,当前帧中包括多个对象,装置400还包括:
候选对象确定模块403,用于获取预先设置的至少一个对象检测模板;分别将当前帧中包括的各对象与对象检测模板进行匹配;确定匹配成功的对象为候选对象。
在一些实施例中,目标对象属于行人对象;对象检测模板包括人体框架模板。候选对象确定模块403还用于分别将当前帧中包括的各对象与人体框架模板进行匹配;根据匹配结果,识别当前帧中包括的行人对象,作为候选对象。
在一些实施例中,候选对象确定模块403还用于针对当前帧中包括的每个对象,对对象进行边缘识别,得到对象的外部轮廓特征;将对象的外部轮廓特征与人体框架模板进行匹配;当匹配成功时,则判定对象为行人对象。
在一些实施例中,目标对象属于行人对象;当前帧中包括多个对象。候选对象确定模块403还用于针对当前帧中的每个对象,从当前帧中截取包括对象的对象区域图;对象占据对象区域图中的主要区域;分别将各个对象区域图输入预先训练的人体识别模型中,输出针对各对象区域图中包括的对象的识别结果;当识别结果表征对象为人体时,则判定对象为候选对象。
在一些实施例中,第一匹配结果包括对象与目标对象之间的第一匹配度;第二匹配结果包括对象与目标对象之间的第二匹配度。特征匹配模块404还用于根据对应于同一对象的第一匹配度和第二匹配度,确定各对象与目标对象之间的最终匹配度。对象识别模块406还用于从各对象所对应的最终匹配度中选取最小的最终匹配度;从当前帧的各对象中,筛选最小的最终匹配度所对应的对象,得到目标对象。
关于对象追踪装置的具体限定可以参见上文中对于对象追踪方法的限定,在此不再赘述。上述对象追踪装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中,也可以以软件形式存储于计算机设备中的存储器中,以便于处理器调用执行以上各个模块对应的操作。
在一些实施例中,提供了一种计算机设备,该计算机设备可以是图1中的服务器120,其内部结构图可以如图6所示。该计算机设备包括通过***总线连接的处理器、存储器和网络接口。存储器包括非易失性存储介质和内存储器。该计算机设备的非易失性存储介质可存储操作***和计算机可读存储介质。该计算机可读存储介质被执行时,可使得处理器执行一种对象追踪方法。该计算机设备的处理器用于提供计算和控制能力,支撑整个计算机设备的运行。该内存储器中可储存有计算机可读存储介质,该计算机可读存储介质被处理器执行时,可使得处理器执行一种对象追踪方法。计算机设备的网络接口用于进行网络 通信。
图7为另一个实施例中计算机设备的内部框图。该计算机设备可以为终端,该计算机设备包括通过***总线连接的处理器、存储器、网络接口、显示屏和输入装置。该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性或易失性存储介质、内存储器。该非易失性或易失性存储介质存储有操作***和计算机可读存储介质。该内存储器为非易失性存储介质中的操作***和计算机可读存储介质的运行提供环境。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读存储介质被处理器执行时以实现一种对象追踪方法。该计算机设备的显示屏可以是液晶显示屏或者电子墨水显示屏,该计算机设备的输入装置可以是显示屏上覆盖的触摸层,也可以是计算机设备外壳上设置的按键、轨迹球或触控板,还可以是外接的键盘、触控板或鼠标等。
本领域技术人员可以理解,图6和图7中示出的结构,仅是与本申请方案相关的部分结构的框图,并不构成对本申请方案所应用于其上的计算机设备的限定,具体的计算机设备可以包括比图中所示更多或更少的部件,或者组合某些部件,或者具有不同的部件布置。
在一些实施例中,提供一种计算机设备,包括存储器和一个或多个处理器,存储器存储有计算机可读存储介质,计算机可读存储介质被处理器执行时,使处理器执行上述对象追踪方法的步骤。对象追踪方法的步骤可以是上述各实施例的对象追踪方法中的步骤。
在一些实施例中,提供了一种计算机可读存储介质,存储有计算机可读存储介质,计算机可读存储介质被处理器执行时,使得处理器执行上述对象追踪方法的步骤。此处对象追踪方法的步骤可以是上述各个实施例的对象追踪方法中的步骤。该计算机可读存储介质可以是非易失性,也可以是易失性的。需要说明的是,本申请各实施例中的“第一”和“第二”等仅用作区分,而并不用于大小、先后、从属等方面的限定。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,是可以通过计算机可读存储介质来指令相关的硬件来完成,所述的计算机可读存储介质可存储于一非易失性或易失性计算机可读取存储介质中,该计算机可读存储介质在执行时,可包括如上述各方法的实施例的流程。本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用,均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限,RAM以多种形式可得,诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
以上实施例的各技术特征可以进行任意的组合,为使描述简洁,未对上述实施例中的各个技术特征所有可能的组合都进行描述,然而,只要这些技术特征的组合不存在矛盾,都应当认为是本说明书记载的范围。以上所述实施例仅表达了本申请的几种实施方式,其描述较为具体和详细,但并不能因此而理解为对发明专利范围的限制。应当指出的是,对于本领域的普通技术人员来说,在不脱离本申请构思的前提下,还可以做出若干变形和改进,这些都属于本申请的保护范围。因此,本申请专利的保护范围应以所附权利要求为准。
Claims (20)
- 一种对象追踪方法,所述方法包括:从多帧图像中选取当前帧;对所述当前帧的前一帧中的目标对象进行特征提取,得到所述目标对象的第一特征;从所述多帧图像中的位于所述当前帧之前的各帧中,分别提取所述目标对象的特征,得到所述目标对象的第二特征;提取所述当前帧中包括的各候选对象的特征;分别将各所述候选对象的特征与所述第一特征进行匹配,得到第一匹配结果,并分别将所述各候选对象的特征与所述第二特征进行匹配,得到第二匹配结果;根据同一所述候选对象对应的所述第一匹配结果和所述第二匹配结果,确定各所述候选对象与所述目标对象之间的最终匹配结果;及按照所述最终匹配结果,从所述当前帧中的各所述候选对象中,识别所述目标对象。
- 根据权利要求1所述的方法,其中,所述当前帧中包括多个对象,在所述提取所述当前帧中包括的各候选对象的特征之前,所述方法还包括:获取预先设置的至少一个对象检测模板;分别将所述当前帧中包括的各对象与所述对象检测模板进行匹配;及确定匹配成功的对象为候选对象。
- 根据权利要求2所述的方法,其中,所述目标对象属于行人对象;所述对象检测模板包括人体框架模板;所述分别将所述当前帧中包括的各对象与所述对象检测模板进行匹配包括:分别将所述当前帧中包括的各对象与所述人体框架模板进行匹配;及所述确定匹配成功的对象为候选对象包括:根据匹配结果,识别所述当前帧中包括的行人对象,作为候选对象。
- 根据权利要求3所述的方法,其中,所述分别将所述当前帧中包括的各对象与所述人体框架模板进行匹配包括:针对所述当前帧中包括的每个对象,对所述对象进行边缘识别,得到所述对象的外部轮廓特征;将所述对象的外部轮廓特征与所述人体框架模板进行匹配;及当匹配成功时,则判定所述对象为行人对象。
- 根据权利要求1所述的方法,其中,所述目标对象属于行人对象;所述当前帧中包括多个对象;在所述提取所述当前帧中包括的各候选对象的特征之前,所述方法还包括:针对所述当前帧中的每个对象,从所述当前帧中截取包括所述对象的对象区域图;所述对象占据所述对象区域图中的主要区域;分别将各个对象区域图输入预先训练的人体识别模型中,输出针对各对象区域图中包括的对象的识别结果;当所述识别结果表征所述对象为人体时,则判定所述对象为候选对象。
- 根据权利要求1至5中任一项所述的方法,其中,所述第一匹配结果包括各所述候选对象与所述目标对象之间的第一匹配度;所述第二匹配结果包括各所述候选对象与所述目标对象之间的第二匹配度;所述根据同一所述候选对象对应的所述第一匹配结果和所述第二匹配结果,确定各所 述候选对象与所述目标对象之间的最终匹配结果包括:根据对应于同一所述候选对象的所述第一匹配度和所述第二匹配度,确定各所述候选对象与所述目标对象之间的最终匹配度;及所述按照所述最终匹配结果,从所述当前帧中的各所述候选对象中,识别所述目标对象包括:从各所述候选对象所对应的最终匹配度中选取最小的最终匹配度;从所述当前帧中的各所述候选对象中,筛选所述最小的最终匹配度所对应的候选对象,作为所述目标对象。
- 一种对象追踪装置,其中,所述装置包括:特征提取模块,用于从多帧图像中选取当前帧;对所述当前帧的前一帧中的目标对象进行特征提取,得到所述目标对象的第一特征;从所述多帧图像中的位于所述当前帧之前的各帧中,分别提取所述目标对象的特征,得到所述目标对象的第二特征;提取所述当前帧中包括的各候选对象的特征;特征匹配模块,用于分别将各所述候选对象的特征与所述第一特征进行匹配,得到第一匹配结果,并分别将所述各候选对象的特征与所述第二特征进行匹配,得到第二匹配结果;根据同一所述候选对象对应的所述第一匹配结果和所述第二匹配结果,确定各所述候选对象与所述目标对象之间的最终匹配结果;及对象识别模块,用于按照所述最终匹配结果,从所述当前帧中的各所述候选对象中,识别所述目标对象。
- 根据权利要求7所述的装置,其特征在于,所述当前帧中包括多个对象,所述装置还包括:候选对象确定模块,用于获取预先设置的至少一个对象检测模板;分别将当前帧中包括的各对象与所述对象检测模板进行匹配;确定匹配成功的对象为候选对象。
- 一种计算机设备,包括存储器和一个或多个处理器,所述存储器存储有计算机可读指令,所述计算机可读指令被所述一个或多个处理器执行时,使得所述一个或多个处理器执行以下步骤:从多帧图像中选取当前帧;对所述当前帧的前一帧中的目标对象进行特征提取,得到所述目标对象的第一特征;从所述多帧图像中的位于所述当前帧之前的各帧中,分别提取所述目标对象的特征,得到所述目标对象的第二特征;提取所述当前帧中包括的各候选对象的特征;分别将各所述候选对象的特征与所述第一特征进行匹配,得到第一匹配结果,并分别将所述各候选对象的特征与所述第二特征进行匹配,得到第二匹配结果;根据同一所述候选对象对应的所述第一匹配结果和所述第二匹配结果,确定各所述候选对象与所述目标对象之间的最终匹配结果;及按照所述最终匹配结果,从所述当前帧中的各所述候选对象中,识别所述目标对象。
- 根据权利要求9所述的计算机设备,其中,所述当前帧中包括多个对象,在所述提取所述当前帧中包括的各候选对象的特征之前,所述处理器执行所述计算机可读指令时还执行以下步骤:获取预先设置的至少一个对象检测模板;分别将所述当前帧中包括的各对象与所述对象检测模板进行匹配;及确定匹配成功的对象为候选对象。
- 根据权利要求10所述的计算机设备,其中,所述目标对象属于行人对象;所述对象检测模板包括人体框架模板;所述分别将所述当前帧中包括的各对象与所述对象检测模板进行匹配包括:分别将所述当前帧中包括的各对象与所述人体框架模板进行匹配;及所述确定匹配成功的对象为候选对象包括:根据匹配结果,识别所述当前帧中包括的行人对象,作为候选对象。
- 根据权利要求11所述的计算机设备,其中,所述分别将所述当前帧中包括的各对象与所述人体框架模板进行匹配包括:针对所述当前帧中包括的每个对象,对所述对象进行边缘识别,得到所述对象的外部轮廓特征;将所述对象的外部轮廓特征与所述人体框架模板进行匹配;及当匹配成功时,则判定所述对象为行人对象。
- 根据权利要求9所述的计算机设备,其中,所述目标对象属于行人对象;所述当前帧中包括多个对象;在所述提取所述当前帧中包括的各候选对象的特征之前,所述处理器执行所述计算机可读指令时还执行以下步骤:针对所述当前帧中的每个对象,从所述当前帧中截取包括所述对象的对象区域图;所述对象占据所述对象区域图中的主要区域;分别将各个对象区域图输入预先训练的人体识别模型中,输出针对各对象区域图中包括的对象的识别结果;及当所述识别结果表征所述对象为人体时,则判定所述对象为候选对象。
- 根据权利要求9至13中任一项所述的计算机设备,其中,第一匹配结果包括各所述候选对象与所述目标对象之间的第一匹配度;所述第二匹配结果包括各所述候选对象与所述目标对象之间的第二匹配度;所述根据同一所述候选对象对应的所述第一匹配结果和所述第二匹配结果,确定各所述候选对象与所述目标对象之间的最终匹配结果包括:根据对应于同一所述候选对象的所述第一匹配度和所述第二匹配度,确定各所述候选对象与所述目标对象之间的最终匹配度;及所述按照所述最终匹配结果,从所述当前帧中的各所述候选对象中,识别所述目标对象包括:从各所述候选对象所对应的最终匹配度中选取最小的最终匹配度;从所述当前帧中的各所述候选对象中,筛选所述最小的最终匹配度所对应的候选对象,作为所述目标对象。
- 一个或多个存储有计算机可读指令的计算机可读存储介质,其中,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器执行以下步骤:从多帧图像中选取当前帧;对所述当前帧的前一帧中的目标对象进行特征提取,得到所述目标对象的第一特征;从所述多帧图像中的位于所述当前帧之前的各帧中,分别提取所述目标对象的特征,得到所述目标对象的第二特征;提取所述当前帧中包括的各候选对象的特征;分别将各所述候选对象的特征与所述第一特征进行匹配,得到第一匹配结果,并分别将所述各候选对象的特征与所述第二特征进行匹配,得到第二匹配结果;根据同一所述候选对象对应的所述第一匹配结果和所述第二匹配结果,确定各所述候选对象与所述目标对象之间的最终匹配结果;及按照所述最终匹配结果,从所述当前帧中的各所述候选对象中,识别所述目标对象。
- 根据权利要求15所述的计算机可读存储介质,其中,所述当前帧中包括多个对象,在所述提取所述当前帧中包括的各候选对象的特征之前,所述计算机可读指令被所述处理器执行时还执行以下步骤:获取预先设置的至少一个对象检测模板;分别将所述当前帧中包括的各对象与所述对象检测模板进行匹配;及确定匹配成功的对象为候选对象。
- 根据权利要求16所述的计算机可读存储介质,其中,所述目标对象属于行人对象;所述对象检测模板包括人体框架模板;所述分别将所述当前帧中包括的各对象与所述对象检测模板进行匹配包括:分别将所述当前帧中包括的各对象与所述人体框架模板进行匹配;及所述确定匹配成功的对象为候选对象包括:根据匹配结果,识别所述当前帧中包括的行人对象,作为候选对象。
- 根据权利要求17所述的计算机可读存储介质,其中,所述分别将所述当前帧中包括的各对象与所述人体框架模板进行匹配包括:针对所述当前帧中包括的每个对象,对所述对象进行边缘识别,得到所述对象的外部轮廓特征;将所述对象的外部轮廓特征与所述人体框架模板进行匹配;及当匹配成功时,则判定所述对象为行人对象。
- 根据权利要求15所述的计算机可读存储介质,其中,所述目标对象属于行人对象;所述当前帧中包括多个对象;在所述提取所述当前帧中包括的各候选对象的特征之前,所述计算机可读指令被所述处理器执行时还执行以下步骤:针对所述当前帧中的每个对象,从所述当前帧中截取包括所述对象的对象区域图;所述对象占据所述对象区域图中的主要区域;分别将各个对象区域图输入预先训练的人体识别模型中,输出针对各对象区域图中包括的对象的识别结果;及当所述识别结果表征所述对象为人体时,则判定所述对象为候选对象。
- 根据权利要求15至19中任一项所述的计算机可读存储介质,其中,所述第一匹配结果包括各所述候选对象与所述目标对象之间的第一匹配度;所述第二匹配结果包括各所述候选对象与所述目标对象之间的第二匹配度;所述根据同一所述候选对象对应的所述第一匹配结果和所述第二匹配结果,确定各所述候选对象与所述目标对象之间的最终匹配结果包括:根据对应于同一所述候选对象的所述第一匹配度和所述第二匹配度,确定各所述候选对象与所述目标对象之间的最终匹配度;及所述按照所述最终匹配结果,从所述当前帧中的各所述候选对象中,识别所述目标对象包括:从各所述候选对象所对应的最终匹配度中选取最小的最终匹配度;从所述当前帧中的各所述候选对象中,筛选所述最小的最终匹配度所对应的候选对象,作为所述目标对象。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910769521.6A CN110660078B (zh) | 2019-08-20 | 2019-08-20 | 对象追踪方法、装置、计算机设备和存储介质 |
CN201910769521.6 | 2019-08-20 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2021031704A1 true WO2021031704A1 (zh) | 2021-02-25 |
Family
ID=69037488
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2020/099170 WO2021031704A1 (zh) | 2019-08-20 | 2020-06-30 | 对象追踪方法、装置、计算机设备和存储介质 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN110660078B (zh) |
WO (1) | WO2021031704A1 (zh) |
Families Citing this family (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110660078B (zh) * | 2019-08-20 | 2024-04-05 | 平安科技(深圳)有限公司 | 对象追踪方法、装置、计算机设备和存储介质 |
CN112578351A (zh) * | 2020-02-28 | 2021-03-30 | 加特兰微电子科技(上海)有限公司 | 目标匹配方法、集成电路、无线电器件及设备 |
CN112183517B (zh) * | 2020-09-22 | 2023-08-11 | 平安科技(深圳)有限公司 | 证卡边缘检测方法、设备及存储介质 |
CN113177967A (zh) * | 2021-03-31 | 2021-07-27 | 千里眼(广州)人工智能科技有限公司 | 视频数据的物体追踪方法、***和存储介质 |
CN116055867B (zh) * | 2022-05-30 | 2023-11-24 | 荣耀终端有限公司 | 一种拍摄方法和电子设备 |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6526156B1 (en) * | 1997-01-10 | 2003-02-25 | Xerox Corporation | Apparatus and method for identifying and tracking objects with view-based representations |
US20070122001A1 (en) * | 2005-11-30 | 2007-05-31 | Microsoft Corporation | Real-time Bayesian 3D pose tracking |
CN105678808A (zh) * | 2016-01-08 | 2016-06-15 | 浙江宇视科技有限公司 | 运动目标跟踪方法及装置 |
CN108470353A (zh) * | 2018-03-01 | 2018-08-31 | 腾讯科技(深圳)有限公司 | 一种目标跟踪方法、装置和存储介质 |
CN110059521A (zh) * | 2018-01-18 | 2019-07-26 | 浙江宇视科技有限公司 | 目标跟踪方法及装置 |
CN110660078A (zh) * | 2019-08-20 | 2020-01-07 | 平安科技(深圳)有限公司 | 对象追踪方法、装置、计算机设备和存储介质 |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR101371275B1 (ko) * | 2012-11-05 | 2014-03-26 | 재단법인대구경북과학기술원 | 스테레오 영상 기반 다중 객체 추적 방법 및 이의 프로그램을 기록한 기록매체 |
CN103049751A (zh) * | 2013-01-24 | 2013-04-17 | 苏州大学 | 一种改进的加权区域匹配高空视频行人识别方法 |
CN105335955B (zh) * | 2014-07-17 | 2018-04-10 | 株式会社理光 | 对象检测方法和对象检测装置 |
CN109816690A (zh) * | 2018-12-25 | 2019-05-28 | 北京飞搜科技有限公司 | 基于深度特征的多目标追踪方法及*** |
-
2019
- 2019-08-20 CN CN201910769521.6A patent/CN110660078B/zh active Active
-
2020
- 2020-06-30 WO PCT/CN2020/099170 patent/WO2021031704A1/zh active Application Filing
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6526156B1 (en) * | 1997-01-10 | 2003-02-25 | Xerox Corporation | Apparatus and method for identifying and tracking objects with view-based representations |
US20070122001A1 (en) * | 2005-11-30 | 2007-05-31 | Microsoft Corporation | Real-time Bayesian 3D pose tracking |
CN105678808A (zh) * | 2016-01-08 | 2016-06-15 | 浙江宇视科技有限公司 | 运动目标跟踪方法及装置 |
CN110059521A (zh) * | 2018-01-18 | 2019-07-26 | 浙江宇视科技有限公司 | 目标跟踪方法及装置 |
CN108470353A (zh) * | 2018-03-01 | 2018-08-31 | 腾讯科技(深圳)有限公司 | 一种目标跟踪方法、装置和存储介质 |
CN110660078A (zh) * | 2019-08-20 | 2020-01-07 | 平安科技(深圳)有限公司 | 对象追踪方法、装置、计算机设备和存储介质 |
Also Published As
Publication number | Publication date |
---|---|
CN110660078A (zh) | 2020-01-07 |
CN110660078B (zh) | 2024-04-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2021031704A1 (zh) | 对象追踪方法、装置、计算机设备和存储介质 | |
CN108470332B (zh) | 一种多目标跟踪方法及装置 | |
US20220392234A1 (en) | Training neural networks for vehicle re-identification | |
Qu et al. | RGBD salient object detection via deep fusion | |
WO2021169111A1 (zh) | 简历筛选方法、装置、计算机设备和存储介质 | |
US20200210786A1 (en) | Computer-executed method and apparatus for assessing vehicle damage | |
WO2019105106A1 (zh) | 图像分类方法及个性化推荐方法、计算机设备及存储介质 | |
CN110209859B (zh) | 地点识别及其模型训练的方法和装置以及电子设备 | |
CN111950329A (zh) | 目标检测及模型训练方法、装置、计算机设备和存储介质 | |
CN109002766B (zh) | 一种表情识别方法及装置 | |
US20170351905A1 (en) | Learning model for salient facial region detection | |
CN112001932B (zh) | 人脸识别方法、装置、计算机设备和存储介质 | |
CN111275060B (zh) | 识别模型更新处理方法、装置、电子设备及存储介质 | |
CN111191568B (zh) | 翻拍图像识别方法、装置、设备及介质 | |
US20120027263A1 (en) | Hand gesture detection | |
WO2019080411A1 (zh) | 电子装置、人脸图像聚类搜索方法和计算机可读存储介质 | |
WO2021114612A1 (zh) | 目标重识别方法、装置、计算机设备和存储介质 | |
WO2020155790A1 (zh) | 理赔信息提取方法和装置、电子设备 | |
US20210056312A1 (en) | Video blocking region selection method and apparatus, electronic device, and system | |
CN110245714B (zh) | 图像识别方法、装置及电子设备 | |
WO2022057309A1 (zh) | 肺部特征识别方法、装置、计算机设备及存储介质 | |
CN110705489B (zh) | 目标识别网络的训练方法、装置、计算机设备和存储介质 | |
US11403875B2 (en) | Processing method of learning face recognition by artificial intelligence module | |
CN113434716A (zh) | 一种跨模态信息检索方法和装置 | |
CN112232397A (zh) | 图像分类模型的知识蒸馏方法、装置和计算机设备 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 20854241 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 20854241 Country of ref document: EP Kind code of ref document: A1 |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 20854241 Country of ref document: EP Kind code of ref document: A1 |