WO2018121287A1 - 目标再识别方法和装置 - Google Patents

目标再识别方法和装置 Download PDF

Info

Publication number
WO2018121287A1
WO2018121287A1 PCT/CN2017/116330 CN2017116330W WO2018121287A1 WO 2018121287 A1 WO2018121287 A1 WO 2018121287A1 CN 2017116330 W CN2017116330 W CN 2017116330W WO 2018121287 A1 WO2018121287 A1 WO 2018121287A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature information
feature
tracking
image
model
Prior art date
Application number
PCT/CN2017/116330
Other languages
English (en)
French (fr)
Inventor
唐矗
孙晓路
Original Assignee
纳恩博(北京)科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 纳恩博(北京)科技有限公司 filed Critical 纳恩博(北京)科技有限公司
Publication of WO2018121287A1 publication Critical patent/WO2018121287A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/40Scenes; Scene-specific elements in video content
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/20Analysis of motion
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/20Image preprocessing
    • G06V10/26Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features
    • G06V10/50Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features
    • G06V10/56Extraction of image or video features relating to colour
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/40Scenes; Scene-specific elements in video content
    • G06V20/46Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames

Definitions

  • the present invention relates to the field of video image processing, and in particular to a method and apparatus for re-identifying a target.
  • Image-based target re-identification generally refers to identifying a given target from different images and videos. Such techniques are generally used in the field of target tracking, content-based image retrieval, and the like.
  • the traditional method is to use the paired image data in different scenes of the same target and the image data of different target pairs, respectively extract the specified features, such as color histograms, as feature vectors, and then learn a similarity by using the metric learning method.
  • the metric function in the application, uses the similarity metric function to calculate the similarity of the two targets, thereby determining whether it is the same target.
  • the tracking-by-detection tracking system can be regarded as a target re-identification process when judging whether the target is the same.
  • a similarity measure function is obtained offline, and in the application, it is directly judged whether the two images are the same. Due to the influence of changes in the appearance of the tracking target caused by the environment and illumination changes during the tracking process, if the target re-identification method is directly applied to the tracking system, two images are used to determine whether the target is the same, and the tracking system is often limited by the environment.
  • the target re-identification problem in visual tracking has a certain difference from the pure target re-recognition.
  • the target in the tracking is re-identified. It is necessary to judge whether the target in the subsequent video frame is the same as the initial setting. A re-identification that finds the same target from an open set, not in a broad sense.
  • the visual tracking maintains an online update template, and uses the template to find the tracking target in a new frame.
  • this method is affected by the appearance change of the tracking target caused by the environment and illumination changes, and the tracking occurs.
  • Errors, and constantly zoom in, difficult to correct, and one of the disadvantages of such methods is that it is difficult to accurately determine whether the tracking target is lost, or it is difficult to retrieve the initial tracking target after the target is lost; in addition, after the target is lost Due to the changes in lighting and environment during the tracking process, the appearance of the target between frames will also change significantly, and it is very difficult to accurately retrieve the target through the appearance.
  • At least some embodiments of the present invention provide a target re-identification method and apparatus to at least solve the technical problem of poor robustness of a re-identification technique for tracking targets in existing tracking technologies.
  • a method for re-identifying a target includes: acquiring a tracking target and an image region of the tracking target; extracting feature information from the image region of the tracking target, and constructing the feature model according to the feature information;
  • the credibility of the tracking result of the frame image determines the tracking state of the tracking target, wherein the credibility of the tracking result is determined by the similarity between the feature information of the preset image region of the current frame and the feature model; and the tracking target is determined according to the tracking state.
  • the feature model is updated according to the tracking result of the current frame image.
  • the feature model is obtained by replacing the original feature information in the preset model with the feature information extracted from the image region of the tracking target.
  • the preset feature information in the feature replaces any one of the feature models with a preset probability to update the feature model.
  • acquiring the feature information of the current frame image and the Pap address of the last updated feature model obtaining the preset probability by using the following formula: Where p is the preset probability, d median is the Pap address of the feature information and the most recently updated feature model, and ⁇ is a preset constant.
  • determining a median value of the feature information of the current frame image and the plurality of Pall's distances of the plurality of feature information in the most recently updated model is the Paging distance of the feature information and the most recently updated model.
  • the background image in the image area of the tracking target is removed; the image area in which the background image is removed is divided into a plurality of images in a preset direction; and the feature information of the plurality of images after the equalization is acquired; The feature information of the images is connected in the order of division, and the feature information of the image of the tracking target is obtained.
  • the feature information is image color feature information, wherein the image color feature information includes: color name information and/or tone information.
  • a target re-identification device comprising: an acquisition module configured to acquire a tracking target and an image region of the tracking target; and a constructing module configured to extract feature information from the image region of the tracking target And constructing a feature model according to the feature information; the determining module is configured to determine a tracking state of the tracking target according to the credibility of the tracking result of the current frame image, wherein the credibility of the tracking result is determined by the current frame The similarity between the feature information of the image preset area and the feature model is determined; the update module is configured to update the feature model according to the tracking result of the current frame image if it is determined that the tracking target is not lost according to the tracking state.
  • the constructing module includes: an initializing submodule configured to replace the original feature information in the preset model by the feature information extracted from the image region of the tracking target to obtain the feature model.
  • the updating module includes: an extracting submodule, configured to perform feature extraction on the image region that obtains the tracking result of the current frame image, and perform normalization processing to obtain corresponding feature information; the first acquiring submodule is set to Obtaining a preset probability; the replacing sub-module is configured to replace any one of the feature models with a preset probability by using preset feature information in the feature information corresponding to the acquired current frame image to update the feature model.
  • the first obtaining sub-module includes: an acquiring unit, configured to acquire feature information of the current frame image and a Pap address distance of the feature model of the latest update; and the calculating unit is configured to obtain the preset probability by using the following formula: Where p is the preset probability, d median is the Pap address of the feature information and the most recently updated feature model, and ⁇ is a preset constant.
  • the obtaining unit includes: a determining subunit, configured to determine a median value of the plurality of Pap s distances of the feature information of the current frame image and the plurality of feature information in the most recently updated model as the feature information and the latest update The Pap sm distance of the model.
  • the constructing module includes: a background removing sub-module configured to remove a background image in the image area of the tracking target; and a dividing sub-module configured to divide the image area from which the background image is removed into a plurality of images in a preset direction;
  • the acquiring sub-module is configured to acquire feature information of the plurality of images after the equalization; and the connecting sub-module is configured to connect the feature information of the plurality of images that are equally divided according to the order of the segmentation, to obtain the feature of the image of the tracking target information.
  • the feature information is image color feature information, wherein the image color feature information includes: color name information and/or tone information.
  • an image region of a tracking target and a tracking target is acquired, feature information is extracted from an image region of the tracking target, and a feature model is constructed according to the feature information, and the reliability of the tracking result of the current frame image is determined.
  • the above scheme constructs a feature model according to the feature information of the image region of the tracking target, and continuously updates the feature model according to different tracking results during the tracking process, and uses the feature model as a tracking model to track, thereby improving the tracking model.
  • the stickiness further solves the technical problem that the re-identification technology of the tracking target in the existing tracking technology is poor in robustness.
  • FIG. 1 is a flow chart of a method of object re-identification according to an embodiment of the present invention
  • FIG. 2 is a schematic diagram of a target re-identification device in accordance with an embodiment of the present invention.
  • an embodiment of a method of object re-identification is provided, it being noted that the steps illustrated in the flowchart of the figures may be performed in a computer system, such as a set of computer-executable instructions. Also, although logical sequences are shown in the flowcharts, in some cases the steps shown or described may be performed in a different order than the ones described herein.
  • FIG. 1 is a flow chart of a method for re-identifying a target according to an embodiment of the present invention. As shown in FIG. 1, the method includes the following steps:
  • Step S102 acquiring a tracking target and an image area of the tracking target.
  • the tracking target may be a target specified by a person or determined by a pedestrian detector
  • the image area of the tracking target may be an artificially included area including a tracking target indicated in a certain frame image of the video, or
  • the pedestrian detector determines the image area in a certain frame of the video.
  • Step S104 extracting feature information from the image region of the tracking target, and constructing the feature model according to the feature information.
  • the extracted feature may be a color feature, an edge feature, or the like of the image. Since the tracking target is usually dynamic in the video, it is difficult to track the shape of the tracking target only, and the accuracy is low. However, for a continuous image in a video, the tracking target's time shape changes continuously with the change of the timestamp, but the features of the image are generally consistent, so the above steps construct the model by the extracted image features.
  • Step S106 Determine a tracking state of the tracking target according to the credibility of the tracking result of the current frame image, wherein the reliability of the tracking result is determined by the similarity between the feature information of the current frame image preset region and the feature model.
  • the foregoing tracking result includes an area and a credibility of the tracking target in the image, and the tracking status of the tracking target may include three states: no loss, low credibility, and loss.
  • the credibility threshold may be set, and if it is determined that the credibility of the tracking result exceeds the preset credibility threshold, it is determined that it is not lost.
  • the image may be determined using the first feature information of the image, or may be determined by using a plurality of feature information of the image.
  • Step S108 In the case that it is determined that the tracking target is not lost according to the tracking state, the feature model is updated according to the tracking result of the current frame image.
  • the foregoing solution may be used to track the process of determining the loss, that is, the process of determining whether the tracking target is lost.
  • the tracking target is a specific person, for example.
  • the tracking task is to track the specific character in multiple videos, and after analyzing each frame of the video, a determination result of determining whether to lose or not is obtained, and if the feature information of the preset area of the current frame image is detected, If the similarity with the feature model is higher than the preset value, it is determined that the image is not lost.
  • the feature model is updated after the tracking result in the current frame image, and the tracking is continued. If the feature information and the feature model of the preset image region of the current frame are detected, If the similarity is lower than the preset value, it is determined that the current frame image has been lost, and the step of retrieving the original tracking target is entered.
  • the foregoing solution may be used to re-recover the original tracking target after the tracking target is lost, for example, the tracking model that can be updated recently when determining that the current frame image has been lost.
  • the previous frame tracking model is used to retrieve the tracking target. Since the previous frame image is in an unfollowed state, the tracking result of the tracking image is updated after obtaining the tracking result of the previous frame image, thereby enabling The feature model of the tracking target is retrieved to be the closest feature model.
  • a specific task may be in a dynamic state in the video, and other environmental information in the video also changes with time, that is, the shape of the specific task is constantly changing in the video, and the video The illumination and environment in the environment are also changing. Therefore, it is very difficult to track or retrieve it by simply tracking the shape of the target. Further, tracking with the feature model of the initially determined tracking target is not accurate. As a result, the feature model of the tracking target introduced by the above scheme can effectively remove the influence of the environment change or the change of the target shape in the process of tracking or retrieving, thereby upgrading the robustness of the tracking model.
  • the above steps of the present application acquire the tracking target and the image region of the tracking target, extract the feature information from the image region of the tracking target, and construct the feature model according to the feature information, and determine the tracking according to the credibility of the tracking result of the current frame image.
  • the tracking state of the target in the case where it is determined that the tracking target is not lost according to the tracking state, the feature model is updated according to the tracking result of the current frame image.
  • the above scheme constructs a feature model according to the feature information of the image region of the tracking target, and continuously updates the feature model according to different tracking results during the tracking process, and uses the feature model as a tracking model to track, thereby improving the tracking model.
  • the stickiness further solves the technical problem that the re-identification technology of the tracking target in the existing tracking technology is poor in robustness.
  • step S102 constructing a feature model according to the feature information, including:
  • Step S1021 The feature model is obtained by replacing the original feature information in the preset model with the feature information extracted from the image region of the tracking target.
  • the corresponding feature model is a color feature model
  • the tracking target may be selected, and the color feature is extracted for the selected target image region, where Using the color histogram as the color feature information, the original model consists of N color histograms, and the initialization stage replaces the N histograms in the original model with the normalized feature histogram h 0 extracted by the selected tracking target image.
  • step S108 updating the feature model according to the tracking result of the current frame image, including:
  • Step S1081 Perform feature extraction on the image region of the obtained tracking result of the current frame image, and perform normalization processing to obtain a corresponding plurality of feature information.
  • step S1083 a preset probability is obtained.
  • Step S1085 Replace any feature information in the feature model with a preset probability by using preset feature information in the feature information corresponding to the acquired current frame image to update the feature model.
  • the above solution of the present application updates the feature model by replacing any one of the feature models with the preset probability by using the feature information of the current frame image, so that the feature model can be changed according to the change of the tracking target.
  • Introducing the latest target features in the new model can effectively preserve the characteristics of each moment in the historical process of target tracking, thus ensuring the diversity of feature information in the model.
  • sexuality thereby improving the robustness of the model, thereby minimizing the impact of the environment, light, etc. on the tracking in the video.
  • step S1083, obtaining a preset probability including:
  • Step S1083a Acquire the Paging distance of the feature information of the current frame image and the feature model of the latest update.
  • step S1083b the preset probability is obtained by the following formula:
  • p is the preset probability
  • d median is the Pap address of the feature information and the most recently updated feature model
  • is a preset constant
  • the preset constant ⁇ is used to determine the probability of controlling the update.
  • step S1083a acquiring the feature information of the current frame image and the Pap address of the last updated feature model includes: determining feature information of the current frame image and the model in the latest update model.
  • the median of the plurality of Pap sigma distances of the plurality of feature information is the Pap address of the feature information and the most recently updated model.
  • the 1-d median obtained by using the above calculation may be used as the color confidence; in the tracking target re-retrieving step, the method may also be used to calculate the candidate target and the tracking target. Similarity is used to select candidate targets; the above scheme takes into account changes in the environment and illumination during long-term target tracking, which leads to changes in the appearance of the target.
  • N feature vectors are used to express the characteristics of the target in different environments; The method of random replacement with a certain probability not only ensures the difference of N feature vectors in the model, but also preserves the historical information of the target, which can effectively improve the robustness of the long-term tracking system.
  • the model does not define specific color features, and the simplest color histogram can be used, or complex feature vector calculation methods can be used.
  • step S104 extracting feature information from the image area of the tracking target, including:
  • Step S1041 removing the background image in the image area of the tracking target.
  • step S1043 the image area from which the background image is removed is divided into a plurality of images in a preset direction.
  • the preset direction may be determined according to a preset tracking target, so as to track the target as a walking person.
  • the preset direction may be a vertical direction.
  • Step S1045 acquiring feature information of the plurality of images after the sharing.
  • step S1047 the feature information of the plurality of divided images is connected in the order of division, and the feature information of the image of the tracking target is obtained.
  • the image features used in the modeling process are color features, specifically, a Color Name histogram, and in addition, a Color Name is calculated.
  • Saliency Segmentation is performed on the image to remove the background interference. Taking the tracking object as the walking character as an example, the pedestrian is mostly in an upright state.
  • the master Before calculating the histogram, the master will be The image after component segmentation is divided into M equal parts in the vertical direction, and each block image is separately statistically histogram; after the M block image histograms are sequentially connected, normalized as color feature information.
  • the feature information is image color feature information, wherein the image color feature information includes: color name information and/or tone information.
  • FIG. 2 is a schematic diagram of a target re-identification device according to an embodiment of the present invention. As shown in FIG. 2, the device includes:
  • the acquisition module 10 is configured to acquire a tracking target and an image area of the tracking target.
  • the tracking target may be a target specified by a person or determined by a pedestrian detector
  • the image area of the tracking target may be an artificially included area including a tracking target indicated in a certain frame image of the video, or by a pedestrian detector.
  • the constructing module 20 is configured to extract feature information from the image region of the tracking target and construct the feature model based on the feature information.
  • the extracted feature may be a color feature, an edge feature, or the like of the image. Since the tracking target is usually dynamic in the video, it is difficult to track the shape of the tracking target only, and the accuracy is low. However, for a continuous image in a video, the tracking target's time shape changes continuously with the change of the timestamp, but the features of the image are generally consistent, so the above steps construct the model by the extracted image features.
  • the determining module 30 is configured to determine a tracking state of the tracking target according to the credibility of the tracking result of the current frame image, wherein the reliability of the tracking result is determined by the similarity between the feature information of the current frame image preset region and the feature model.
  • the foregoing tracking result includes an area and a credibility of the tracking target in the image, and the tracking target is followed by The trace status can include three states: no loss, low confidence, and loss.
  • the credibility threshold may be set, and if it is determined that the credibility of the tracking result exceeds the preset credibility threshold, it is determined that it is not lost.
  • the image may be determined using the first feature information of the image, or may be determined by using a plurality of feature information of the image.
  • the update module 40 is configured to update the feature model according to the tracking result of the current frame image if it is determined that the tracking target is not lost according to the tracking state.
  • a specific task may be in a dynamic state in the video, and other environmental information in the video also changes with time, that is, the shape of the specific task is constantly changing in the video, and in the video. The lighting and environment are also changing. Therefore, it is very difficult to track or retrieve the target by simply tracking the shape of the target. Further, tracking with the feature model of the initially determined tracking target cannot be accurately obtained. As a result, therefore, the feature model of the tracking target introduced by the above scheme can effectively remove the influence of the change of the environment or the change of the shape of the target in the process of tracking or retrieving, thereby upgrading the robustness of the tracking model.
  • the above solution of the present application acquires the tracking target and the image region of the tracking target through the acquiring module, extracts the feature information from the image region of the tracking target through the constructing module, and constructs the feature model according to the feature information, and determines the module according to the current frame image.
  • the credibility of the tracking result determines the tracking state of the tracking target
  • the update module updates the feature model according to the tracking result of the current frame image in the case that the tracking target is not lost according to the tracking state.
  • the above scheme constructs a feature model according to the feature information of the image region of the tracking target, and continuously updates the feature model according to different tracking results during the tracking process, and uses the feature model as a tracking model to track, thereby improving the tracking model.
  • the stickiness further solves the technical problem that the re-identification technology of the tracking target in the existing tracking technology is poor in robustness.
  • the foregoing constructing module includes:
  • the initialization sub-module is configured to replace the original feature information in the preset model by the feature information extracted from the image region of the tracking target to obtain the feature model.
  • the foregoing update module includes:
  • Extracting a sub-module performing feature extraction on an image region that obtains a tracking result of the current frame image, and performing normalization processing to obtain a corresponding plurality of feature information
  • the first obtaining submodule is configured to acquire a preset probability
  • the replacement sub-module is configured to replace any one of the feature models with a preset probability by using preset feature information in the feature information corresponding to the acquired current frame image to update the feature model.
  • the foregoing first obtaining submodule includes:
  • An obtaining unit configured to acquire a plurality of feature information of the current frame image and a Pap address of the most recently updated feature model
  • the calculation unit is set to obtain the preset probability by the following formula:
  • p is the preset probability
  • d median is the Pap address of the feature information and the most recently updated feature model
  • is a preset constant
  • the acquiring unit includes:
  • the determining subunit is set to determine that the median information of the feature information of the current frame image and the plurality of Pap sigma distances of the most recently updated model is the Paging distance of the plurality of feature information and the most recently updated model.
  • the foregoing constructing module includes:
  • a dividing sub-module configured to divide the image area from which the background image is removed into a plurality of images in a preset direction
  • a second obtaining submodule configured to acquire feature information of the plurality of images after the equalization
  • the connection submodule is configured to connect the feature information of the plurality of divided images in the order of division to obtain the feature information of the image of the tracking target.
  • the feature information is image color feature information, wherein the image color feature information includes: color name information and/or tone information.
  • a storage medium wherein the storage medium comprises a stored program, wherein the device in which the storage medium is located is controlled to execute the target re-identification method when the program is running.
  • the above storage medium may include, but is not limited to, a U disk, a read only memory (ROM), a random access memory (RAM), a removable hard disk, a magnetic disk, or an optical disk, and the like, which can store program codes.
  • a processor configured to execute a program, wherein the target re-identification method is executed when the program is running.
  • the above processor may include, but is not limited to, a processing device such as a microprocessor (MCU) or a programmable logic device (FPGA).
  • MCU microprocessor
  • FPGA programmable logic device
  • the disclosed technical contents may be implemented in other manners.
  • the device embodiments described above are only schematic.
  • the division of the unit may be a logical function division.
  • there may be another division manner for example, multiple units or components may be combined or may be Integrate into another system, or some features can be ignored or not executed.
  • the mutual coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interface, unit or module, and may be electrical or otherwise.
  • the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of the embodiment.
  • each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.
  • the above integrated unit can be implemented in the form of hardware or in the form of a software functional unit.
  • the integrated unit if implemented in the form of a software functional unit and sold or used as a standalone product, may be stored in a computer readable storage medium.
  • the technical solution of the present invention which is essential or contributes to the prior art, or all or part of the technical solution, may be embodied in the form of a software product stored in a storage medium.
  • a number of instructions are included to cause a computer device (which may be a personal computer, server or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention.
  • At least some embodiments of the present invention provide a target re-identification method and apparatus having the following beneficial effects: constructing a feature model according to feature information of an image region of a tracking target, and continuously according to different tracking results during the tracking process.
  • the feature model is updated, and the feature model is used as the tracking model to track, thus improving the robustness of the tracking model.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

一种目标再识别方法和装置。其中,该方法包括:获取跟踪目标以及跟踪目标的图像区域(S102);从跟踪目标的图像区域中提取特征信息,并根据特征信息构造特征模型(S104);根据当前帧图像的跟踪结果的可信度确定跟踪目标的跟踪状态(S106),其中,跟踪结果的可信度由当前帧图像预设区域的特征信息与特征模型的相似度确定;在根据跟踪状态确定跟踪目标未丢失的情况下,根据当前帧图像的跟踪结果更新特征模型(S108)。解决了现有跟踪技术中跟踪目标的再识别技术鲁棒性差的技术问题。

Description

目标再识别方法和装置 技术领域
本发明涉及视频图像处理领域,具体而言,涉及一种目标再识别方法和装置。
背景技术
基于图像的目标再识别一般是指从不同的图像、视频中识别出给定的目标,该类技术一般用于跨场景的目标跟踪、基于内容的图像检索等领域。传统的做法是使用同一目标不同场景下成对的图像数据以及不同目标成对的图像数据,分别提取指定的特征,如颜色直方图,作为特征向量,然后利用度量学习的方法学习到一个相似度度量函数,在应用中,利用该相似度度量函数计算两个目标的相似度,进而判别是否是同一个目标。
近来,随着深度学习的兴起,也有利用卷积神经网络进行目标再识别的出现,思路与传统方法类似,不同之处在于,不需要人为指定特征,表征目标相似性与差异性的特征以及衡量特征相似性的函数都由卷积神经网络自动学习得到,在应用中,将学习到的卷积网络模型应用于两幅图像,进而判断目标是否是同一个。
视觉跟踪的本质是在不同的帧间找到同一个目标的位置,一般而言tracking-by-detection的跟踪***,在判断目标是否是同一个时候,即可看做是一个目标再识别的过程。但不管是传统的目标再识别方法还是基于深度学习的方法,都是离线获得一个相似度度量函数,在应用中,直接判断两幅图像是否是同一个。由于跟踪过程中环境、光照变化引起的跟踪目标外观变化的影响,如果将这种目标再识别方式直接应用于跟踪***,用两张图片判断目标是否是同一个,跟踪***往往会受限于环境的变换;此外,视觉跟踪中目标再识别问题与纯粹的目标再识别也具有一定的差异性,跟踪中的目标再识别,需要判断的是后续视频帧中的目标与初始设定的是否是同一个,而非广泛意义上的从一个开集中找到相同目标的再识别,。
通常视觉跟踪会维持一个在线更新模板,并利用该模板在新的一帧中找到跟踪目标,在长时间跟踪过程中,这种方法会受到环境、光照变化引起的跟踪目标外观变化影响,出现跟踪错误,并不断放大,难以纠正,而此类方法的一个缺点在于本身很难准确地判断跟踪目标是否丢失,或是在目标跟丢之后,难以找回初始的跟踪目标;此外,在目标丢失后,由于跟踪过程中光照、环境等变化,相隔帧间目标的外观也会发生明显变化,通过外观准确地找回目标非常困难。
针对现有跟踪技术中跟踪目标的再识别技术鲁棒性差的问题,目前尚未提出有效 的解决方案。
发明内容
本发明至少部分实施例提供了一种目标再识别方法和装置,以至少解决现有跟踪技术中跟踪目标的再识别技术鲁棒性差的技术问题。
根据本发明其中一实施例,提供了一种目标再识别方法,包括:获取跟踪目标以及跟踪目标的图像区域;从跟踪目标的图像区域中提取特征信息,并根据特征信息构造特征模型;根据当前帧图像的跟踪结果的可信度确定跟踪目标的跟踪状态,其中,跟踪结果的可信度由当前帧图像预设区域的特征信息与特征模型的相似度确定;在根据跟踪状态确定跟踪目标未丢失的情况下,根据当前帧图像的跟踪结果更新特征模型。
可选地,通过从跟踪目标的图像区域中提取的特征信息替换预设模型中的原始特征信息,得到特征模型。
可选地,对获取的当前帧图像的跟踪结果的图像区域进行特征提取,并进行归一化处理,得到对应的多个特征信息;获取预设概率;通过获取的当前帧图像对应的特征信息中的预设特征信息以预设概率替换特征模型中的任意一个特征信息,以更新特征模型。
可选地,获取当前帧图像的特征信息与最近一次更新的特征模型的巴氏距离;通过如下公式获取预设概率:
Figure PCTCN2017116330-appb-000001
其中,p为预设概率,dmedian为特征信息与最近一次更新的特征模型的巴氏距离,σ为预设常数。
可选地,确定当前帧图像的特征信息与最近一次更新的模型中多个特征信息的多个巴氏距离的中值为特征信息与最近一次更新的模型的巴氏距离。
可选地,去除跟踪目标的图像区域中的背景图像;将去除背景图像的图像区域沿预设方向分割为多个图像;获取均分后的多个图像的特征信息;将均分后的多个图像的特征信息按照分割的顺序进行连接,得到跟踪目标的图像的特征信息。
可选地,特征信息为图像颜色特征信息,其中,图像颜色特征信息包括:颜色名称信息和/色调信息。
根据本发明其中一实施例,还提供了一种目标再识别装置,包括:获取模块,设置为获取跟踪目标以及跟踪目标的图像区域;构造模块,设置为从跟踪目标的图像区域中提取特征信息,并根据特征信息构造特征模型;确定模块,设置为根据当前帧图像的跟踪结果的可信度确定跟踪目标的跟踪状态,其中,跟踪结果的可信度由当前帧 图像预设区域的特征信息与特征模型的相似度确定;更新模块,设置为在根据跟踪状态确定跟踪目标未丢失的情况下,根据当前帧图像的跟踪结果更新特征模型。
可选地,构造模块包括:初始化子模块,设置为通过从跟踪目标的图像区域中提取的特征信息替换预设模型中的原始特征信息,得到特征模型。
可选地,更新模块包括:提取子模块,设置为对获取当前帧图像的跟踪结果的图像区域进行特征提取,并进行归一化处理,得到对应的特征信息;第一获取子模块,设置为获取预设概率;替换子模块,设置为通过获取的当前帧图像对应的特征信息中的预设特征信息以预设概率替换特征模型中的任意一个特征信息,以更新特征模型。
可选地,第一获取子模块包括:获取单元,设置为获取当前帧图像的特征信息与最近一次更新的特征模型的巴氏距离;计算单元,设置为通过如下公式获取预设概率:
Figure PCTCN2017116330-appb-000002
其中,p为预设概率,dmedian为特征信息与最近一次更新的特征模型的巴氏距离,σ为预设常数。
可选地,获取单元包括:确定子单元,设置为确定当前帧图像的特征信息与最近一次更新的模型中的多个特征信息的多个巴氏距离的中值为特征信息与最近一次更新的模型的巴氏距离。
可选地,构造模块包括:去除背景子模块,设置为去除跟踪目标的图像区域中的背景图像;分割子模块,设置为将去除背景图像的图像区域沿预设方向分割为多个图像;第二获取子模块,设置为获取均分后的多个图像的特征信息;连接子模块,设置为将均分后的多个图像的特征信息按照分割的顺序进行连接,得到跟踪目标的图像的特征信息。
可选地,特征信息为图像颜色特征信息,其中,图像颜色特征信息包括:颜色名称信息和/色调信息。
在本发明至少部分实施例中,获取跟踪目标以及跟踪目标的图像区域,从跟踪目标的图像区域中提取特征信息,并根据特征信息构造特征模型,根据当前帧图像的跟踪结果的可信度确定跟踪目标的跟踪状态,在根据跟踪状态确定跟踪目标未丢失的情况下,根据当前帧图像的跟踪结果更新特征模型。上述方案根据跟踪目标的图像区域的特征信息构建特征模型,并在跟踪过程中不断的根据不同的跟踪结果对特征模型进行更新,以特征模型作为跟踪模型来进行跟踪,从而提高了跟踪模型的鲁棒性,进而解决了现有跟踪技术中跟踪目标的再识别技术鲁棒性差的技术问题。
附图说明
此处所说明的附图用来提供对本发明的进一步理解,构成本申请的一部分,本发明的示意性实施例及其说明用于解释本发明,并不构成对本发明的不当限定。在附图中:
图1是根据本发明其中一实施例的目标再识别方法的流程图;以及
图2是根据本发明其中一实施例的目标再识别装置的示意图。
具体实施方式
为了使本技术领域的人员更好地理解本发明方案,下面将结合本发明实施例中的附图,对本发明实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本发明一部分的实施例,而不是全部的实施例。基于本发明中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都应当属于本发明保护的范围。
需要说明的是,本发明的说明书和权利要求书及上述附图中的术语“第一”、“第二”等是用于区别类似的对象,而不必用于描述特定的顺序或先后次序。应该理解这样使用的数据在适当情况下可以互换,以便这里描述的本发明的实施例能够以除了在这里图示或描述的那些以外的顺序实施。此外,术语“包括”和“具有”以及他们的任何变形,意图在于覆盖不排他的包含,例如,包含了一系列步骤或单元的过程、方法、***、产品或设备不必限于清楚地列出的那些步骤或单元,而是可包括没有清楚地列出的或对于这些过程、方法、产品或设备固有的其它步骤或单元。
实施例1
根据本发明其中一实施例,提供了一种目标再识别方法的实施例,需要说明的是,在附图的流程图示出的步骤可以在诸如一组计算机可执行指令的计算机***中执行,并且,虽然在流程图中示出了逻辑顺序,但是在某些情况下,可以以不同于此处的顺序执行所示出或描述的步骤。
图1是根据本发明其中一实施例的目标再识别方法的流程图,如图1所示,该方法包括如下步骤:
步骤S102,获取跟踪目标以及跟踪目标的图像区域。
具体的,上述跟踪目标可以是人为指定或通过行人检测器确定的目标,跟踪目标的图像区域可以是人为的在视频的某一帧图像中指出的包含跟踪目标的区域,或通过 行人检测器在视频的某一帧确定的图像区域。
步骤S104,从跟踪目标的图像区域中提取特征信息,并根据特征信息构造特征模型。
具体的,上述提取的特征可以是图像的颜色特征、边缘特征等,由于跟踪目标在视频中通常是动态的,因此仅以跟踪目标的外形为模型跟踪具有一定的难度,且准确度较低,但通常对于视频中的连续图像来说,跟踪目标及时外形随着时间戳的变动而不断变化,但图像的特征通常保持一致,因此上述步骤通过提取的图像特征来构造模型。
步骤S106,根据当前帧图像的跟踪结果的可信度确定跟踪目标的跟踪状态,其中,跟踪结果的可信度由当前帧图像预设区域的特征信息与特征模型的相似度确定。
具体的,上述跟踪结果包括图像中跟踪目标的区域和可信度,上述跟踪目标的跟踪状态可以包括未丢失、可信度低以及丢失三种状态。在一种可选的实施例中,可以设置可信度阈值,如果确定跟踪结果的可信度超过预设的可信度阈值,则确定未丢失。在根据当前图像预设区域的特征信息与特征模型的相似度确定跟踪结果的可信度时,可以使用图像第一种特征信息确定,也可以使用图像的多种特征信息进行信息融合后确定。
步骤S108,在根据跟踪状态确定跟踪目标未丢失的情况下,根据当前帧图像的跟踪结果更新特征模型。
在一种可选的实施例中,上述方案可以用于跟踪判丢的过程,即判断跟踪目标是否丢失的过程,在一种可选的实施例中,以跟踪目标为一个特定人物为例,跟踪任务为在多个视频中跟踪该特定人物,在对视频的每一帧图像进行分析后,都能过得到一个确定是否跟丢的判断结果,如果检测到当前帧图像预设区域的特征信息与特征模型的相似度高于预设值,则确定未跟丢,以当前帧图像中的跟踪结果来更新特征模型后继续进行跟踪,如果检测到当前帧图像预设区域的特征信息与特征模型的相似度低于预设值,则确定当前帧图像已跟丢,并进入找回原跟踪目标的步骤。
在另一种可选的实施例中,上述方案可以用于在跟踪目标丢失后重新找回原跟踪目标的过程,例如,在确定当前帧图像已跟丢的情况下,可以最近更新的跟踪模型,即上一帧跟踪模型为依据来找回跟踪目标,由于上一帧图像为未跟丢的状态,因此得到上一帧图像的跟踪结果后还更新了跟踪目标的特征模型,从而使得用于找回跟踪目标的特征模型为最接近的特征模型。
此处需要说明的是,由于特定任务在视频中可能出于动态的状态,且视频中其他环境信息也随着时间而改变,也就是说特定任务的外形在视频为不断变化的,且视频 中的光照、环境也是变化的,因此单纯的通过跟踪目标的外形来进行跟踪或再找回是十分困难的,进一步的,一直使用最初确定的跟踪目标的特征模型来进行跟踪也并不能得到准确的结果,因此,上述方案引入的跟踪目标的特征模型能够有效的在跟踪或再找回的过程中去除环境的变化或跟踪目标外形的变化的影响,从而升级了跟踪模型的鲁棒性。
由上可知,本申请上述步骤获取跟踪目标以及跟踪目标的图像区域,从跟踪目标的图像区域中提取特征信息,并根据特征信息构造特征模型,根据当前帧图像的跟踪结果的可信度确定跟踪目标的跟踪状态,在根据跟踪状态确定跟踪目标未丢失的情况下,根据当前帧图像的跟踪结果更新特征模型。上述方案根据跟踪目标的图像区域的特征信息构建特征模型,并在跟踪过程中不断的根据不同的跟踪结果对特征模型进行更新,以特征模型作为跟踪模型来进行跟踪,从而提高了跟踪模型的鲁棒性,进而解决了现有跟踪技术中跟踪目标的再识别技术鲁棒性差的技术问题。
可选的,根据本申请上述实施例,步骤S102,根据特征信息构造特征模型,包括:
步骤S1021,通过从跟踪目标的图像区域中提取的特征信息替换预设模型中的原始特征信息,得到特征模型。
在一种可选的实施例中,以特征信息为图像的颜色特征信息为例,对应的特征模型即为颜色特征模型,可以选定跟踪目标,对选定的目标图像区域提取颜色特征,这里使用颜色直方图作为颜色特征信息,原始的模型由N个颜色直方图组成,初始化阶段用选定的跟踪目标图像提取的归一化后的特征直方图h0来替换原始模型中的N个直方图。
可选的,根据本申请上述实施例,步骤S108,根据当前帧图像的跟踪结果更新特征模型,包括:
步骤S1081,对获取到的当前帧图像的跟踪结果的图像区域进行特征提取,并进行归一化处理,得到对应的多个特征信息。
步骤S1083,获取预设概率。
步骤S1085,通过获取的当前帧图像对应的特征信息中的预设特征信息以预设概率替换特征模型中的任意一个特征信息,以更新特征模型。
由上可知,本申请上述方案通过当前帧图像的特征信息以预设概率替换特征模型中的任意一个信息来更新特征模型,以使特征模型能够跟随跟踪目标的变化而变化,一方面,能够确保在新的模型中引入最新的目标特征,另一方面,引入的随机性能够有效地保留目标跟踪的历史过程中各个时刻的特征,从而保证模型中特征信息的多样 性,从而提高模型的鲁棒性,进而尽可能的减小视频中环境、光线等对跟踪的影响。
可选的,根据本申请上述实施例,步骤S1083,获取预设概率,包括:
步骤S1083a,获取当前帧图像的特征信息与最近一次更新的特征模型的巴氏距离。
步骤S1083b,通过如下公式获取预设概率:
Figure PCTCN2017116330-appb-000003
其中,p为预设概率,dmedian为特征信息与最近一次更新的特征模型的巴氏距离,σ为预设常数。
具体的,上述预设常数σ用于确定控制更新的概率。
可选的,根据本申请上述实施例,步骤S1083a,获取当前帧图像的特征信息与最近一次更新的特征模型的巴氏距离,包括:确定当前帧图像的特征信息与最近一次更新的模型中的多个特征信息的多个巴氏距离的中值为特征信息与最近一次更新的模型的巴氏距离。
在一种可选的实施例中,以当前帧图像的特征信息为ht为例,可以逐个计算ht与N个直方图的巴氏距离得到di,i=1,2,3...,N,将di进行升序排列,取中值dmedian作为ht与该模型的距离。
可选地,判断跟踪目标是否丢失的过程中,可以使用上述计算得到的1-dmedian作为颜色置信度;在跟踪目标重找回步骤中,也可以使用此方法,计算候选目标与跟踪目标的相似度来选择候选目标;上述方案考虑到长时间目标跟踪过程中的环境、光照等变化会带来目标外观的变化,采用N个特征向量来表达目标在不同环境下的特征;在更新时,采用以一定概率随机替换的方式,既保证了模型中N个特征向量的差异性,又保留了目标的历史信息,可以有效地提升长时间跟踪***的鲁棒性。此外,该模型不限定具体的颜色特征,可以使用最简单的颜色直方图,也可以使用复杂的特征向量计算方法。
可选的,根据本申请上述实施例,步骤S104,从跟踪目标的图像区域中提取特征信息,包括:
步骤S1041,去除跟踪目标的图像区域中的背景图像。
步骤S1043,将去除背景图像的图像区域沿预设方向分割为多个图像。
具体的,上述预设方向可以根据预设跟踪目标来确定,以跟踪目标为行走的人物 为例,由于人物行走通常为直立状,因此预设方向可以为竖直方向。
步骤S1045,获取均分后的多个图像的特征信息。
步骤S1047,将均分后的多个图像的特征信息按照分割的顺序进行连接,得到跟踪目标的图像的特征信息。
在一种可选的实施例中,如上述实施例所示,在建模的过程中使用的图像特征为颜色特征,具体的,为Color Name(颜色名)直方图,此外,在计算Color Name直方图之前,先对图像进行Saliency Segmentation(主成分分割),以去除背景的干扰;以跟踪对象为行走中的人物为例,针对行人多数是直立状态的特点,在计算直方图前,将主成分分割后的图像沿竖直方向均分为M等份,每一块图像单独统计直方图;将M块图像直方图顺序连接后,归一化做为颜色特征信息。
可选的,根据本申请上述实施例,特征信息为图像颜色特征信息,其中,图像颜色特征信息包括:颜色名称信息和/色调信息。
实施例2
根据本发明其中一实施例,提供了一种目标再识别装置的实施例,图2是根据本发明其中一实施例的目标再识别装置的示意图,如图2所示,该装置包括:
获取模块10,设置为获取跟踪目标以及跟踪目标的图像区域。
具体的,上述跟踪目标可以是人为指定或通过行人检测器确定的目标,跟踪目标的图像区域可以是人为的在视频的某一帧图像中指出的包含跟踪目标的区域,或通过行人检测器在视频的某一帧确定的图像区域。
构造模块20,设置为从跟踪目标的图像区域中提取特征信息,并根据特征信息构造特征模型。
具体的,上述提取的特征可以是图像的颜色特征、边缘特征等,由于跟踪目标在视频中通常是动态的,因此仅以跟踪目标的外形为模型跟踪具有一定的难度,且准确度较低,但通常对于视频中的连续图像来说,跟踪目标及时外形随着时间戳的变动而不断变化,但图像的特征通常保持一致,因此上述步骤通过提取的图像特征来构造模型。
确定模块30,设置为根据当前帧图像的跟踪结果的可信度确定跟踪目标的跟踪状态,其中,跟踪结果的可信度由当前帧图像预设区域的特征信息与特征模型的相似度确定。
具体的,上述跟踪结果包括图像中跟踪目标的区域和可信度,上述跟踪目标的跟 踪状态可以包括未丢失、可信度低以及丢失三种状态。在一种可选的实施例中,可以设置可信度阈值,如果确定跟踪结果的可信度超过预设的可信度阈值,则确定未丢失。在根据当前图像预设区域的特征信息与特征模型的相似度确定跟踪结果的可信度时,可以使用图像第一种特征信息确定,也可以使用图像的多种特征信息进行信息融合后确定。
更新模块40,设置为在根据跟踪状态确定跟踪目标未丢失的情况下,根据当前帧图像的跟踪结果更新特征模型。
此处需要说明的是,由于特定任务在视频中可能出于动态的状态,且视频中其他环境信息也随着时间而改变,也就是说特定任务的外形在视频为不断变化的,且视频中的光照、环境也是变化的,因此单纯的通过跟踪目标的外形来进行跟踪或再找回是十分困难的,进一步的,一直使用最初确定的跟踪目标的特征模型来进行跟踪也并不能得到准确的结果,因此,上述方案引入的跟踪目标的特征模型能够有效的在跟踪或再找回的过程中去除环境的变化或跟踪目标外形的变化的影响,从而升级了跟踪模型的鲁棒性。
由上可知,本申请上述方案通过获取模块获取跟踪目标以及跟踪目标的图像区域,通过构造模块从跟踪目标的图像区域中提取特征信息,并根据特征信息构造特征模型,通过确定模块根据当前帧图像的跟踪结果的可信度确定跟踪目标的跟踪状态,通过更新模块在根据跟踪状态确定跟踪目标未丢失的情况下,根据当前帧图像的跟踪结果更新特征模型。上述方案根据跟踪目标的图像区域的特征信息构建特征模型,并在跟踪过程中不断的根据不同的跟踪结果对特征模型进行更新,以特征模型作为跟踪模型来进行跟踪,从而提高了跟踪模型的鲁棒性,进而解决了现有跟踪技术中跟踪目标的再识别技术鲁棒性差的技术问题。
可选的,根据本申请上述实施例,上述构造模块包括:
初始化子模块,设置为通过从跟踪目标的图像区域中提取的特征信息替换预设模型中的原始特征信息,得到特征模型。
可选的,根据本申请上述实施例,上述更新模块包括:
提取子模块,设置为对获取当前帧图像的跟踪结果的图像区域进行特征提取,并进行归一化处理,得到对应的多个特征信息;
第一获取子模块,设置为获取预设概率;
替换子模块,设置为通过获取的当前帧图像对应的特征信息中的预设特征信息以预设概率替换特征模型中的任意一个特征信息,以更新特征模型。
可选的,根据本申请上述实施例,上述第一获取子模块包括:
获取单元,设置为获取当前帧图像的多个特征信息与最近一次更新的特征模型的巴氏距离;
计算单元,设置为通过如下公式获取预设概率:
Figure PCTCN2017116330-appb-000004
其中,p为预设概率,dmedian为特征信息与最近一次更新的特征模型的巴氏距离,σ为预设常数。
可选的,根据本申请上述实施例,上述获取单元包括:
确定子单元,设置为确定当前帧图像的特征信息与最近一次更新的模型的多个巴氏距离的中值为多个特征信息与最近一次更新的模型的巴氏距离。
可选的,根据本申请上述实施例,上述构造模块包括:
去除背景子模块,设置为去除跟踪目标的图像区域中的背景图像;
分割子模块,设置为将去除背景图像的图像区域沿预设方向分割为多个图像;
第二获取子模块,设置为获取均分后的多个图像的特征信息;
连接子模块,设置为将均分后的多个图像的特征信息按照分割的顺序进行连接,得到跟踪目标的图像的特征信息。
可选的,根据本申请上述实施例,上述特征信息为图像颜色特征信息,其中,图像颜色特征信息包括:颜色名称信息和/色调信息。
根据本发明其中一实施例,还提供了一种存储介质,存储介质包括存储的程序,其中,在程序运行时控制存储介质所在设备执行上述目标再识别方法。上述存储介质可以包括但不限于:U盘、只读存储器(ROM)、随机存取存储器(RAM)、移动硬盘、磁碟或者光盘等各种可以存储程序代码的介质。
根据本发明其中一实施例,还提供了一种处理器,处理器用于运行程序,其中,程序运行时执行上述目标再识别方法。上述处理器可以包括但不限于:微处理器(MCU)或可编程逻辑器件(FPGA)等的处理装置。
上述本发明实施例序号仅仅为了描述,不代表实施例的优劣。
在本发明的上述实施例中,对各个实施例的描述都各有侧重,某个实施例中没有 详述的部分,可以参见其他实施例的相关描述。
在本申请所提供的几个实施例中,应该理解到,所揭露的技术内容,可通过其它的方式实现。其中,以上所描述的装置实施例仅仅是示意性的,例如所述单元的划分,可以为一种逻辑功能划分,实际实现时可以有另外的划分方式,例如多个单元或组件可以结合或者可以集成到另一个***,或一些特征可以忽略,或不执行。另一点,所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口,单元或模块的间接耦合或通信连接,可以是电性或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的,作为单元显示的部件可以是或者也可以不是物理单元,即可以位于一个地方,或者也可以分布到多个单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外,在本发明各个实施例中的各功能单元可以集成在一个处理单元中,也可以是各个单元单独物理存在,也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现,也可以采用软件功能单元的形式实现。
所述集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时,可以存储在一个计算机可读取存储介质中。基于这样的理解,本发明的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来,该计算机软件产品存储在一个存储介质中,包括若干指令用以使得一台计算机设备(可为个人计算机、服务器或者网络设备等)执行本发明各个实施例所述方法的全部或部分步骤。
以上所述仅是本发明的优选实施方式,应当指出,对于本技术领域的普通技术人员来说,在不脱离本发明原理的前提下,还可以做出若干改进和润饰,这些改进和润饰也应视为本发明的保护范围。
工业实用性
如上所述,本发明至少部分实施例提供的一种目标再识别方法和装置具有以下有益效果:根据跟踪目标的图像区域的特征信息构建特征模型,并在跟踪过程中不断的根据不同的跟踪结果对特征模型进行更新,以特征模型作为跟踪模型来进行跟踪,从而提高了跟踪模型的鲁棒性。

Claims (16)

  1. 一种目标再识别方法,包括:
    获取跟踪目标以及所述跟踪目标的图像区域;
    从所述跟踪目标的图像区域中提取特征信息,并根据所述特征信息构造特征模型;
    根据当前帧图像的跟踪结果的可信度确定跟踪目标的跟踪状态,其中,所述跟踪结果的可信度由所述当前帧图像预设区域的特征信息与所述特征模型的相似度确定;
    在根据所述跟踪状态确定所述跟踪目标未丢失的情况下,根据所述当前帧图像的跟踪结果更新所述特征模型。
  2. 根据权利要求1所述的方法,其中,根据所述特征信息构造特征模型,包括:
    通过从所述跟踪目标的图像区域中提取的特征信息替换预设模型中的原始特征信息,得到所述特征模型。
  3. 根据权利要求2所述的方法,其中,根据所述当前帧图像的跟踪结果更新所述特征模型,包括:
    对获取的当前帧图像的跟踪结果的图像区域进行特征提取,并进行归一化处理,得到对应的特征信息;
    获取预设概率;
    通过获取的当前帧图像对应的特征信息中的预设特征信息以预设概率替换所述特征模型中的任意一个特征信息,以更新所述特征模型。
  4. 根据权利要求3所述的方法,其中,获取预设概率,包括:
    获取所述当前帧图像的特征信息与最近一次更新的特征模型的巴氏距离;
    通过如下公式获取所述预设概率:
    Figure PCTCN2017116330-appb-100001
    其中,所述p为所述预设概率,所述dmedian为所述特征信息与所述最近一次更新的特征模型的巴氏距离,所述σ为预设常数。
  5. 根据权利要求4所述的方法,其中,获取所述当前帧图像的特征信息与所述最近一次更新的特征模型的巴氏距离,包括:
    确定所述当前帧图像的特征信息与所述最近一次更新的模型中的多个特征信息的多个巴氏距离的中值为所述特征信息与所述最近一次更新的模型的巴氏距离。
  6. 根据权利要求1至5中任意一项所述的方法,其中,从所述跟踪目标的图像区域中提取特征信息,包括:
    去除所述跟踪目标的图像区域中的背景图像;
    将去除所述背景图像的所述图像区域沿预设方向分割为多个图像;
    获取均分后的所述多个图像的特征信息;
    将均分后的所述多个图像的特征信息按照分割的顺序进行连接,得到所述跟踪目标的图像的特征信息。
  7. 根据权利要求1至5中任意一项所述的方法,其中,所述特征信息为图像颜色特征信息,其中,所述图像颜色特征信息包括:颜色名称信息和/或色调信息。
  8. 一种目标再识别装置,包括:
    获取模块,设置为获取跟踪目标以及所述跟踪目标的图像区域;
    构造模块,设置为从所述跟踪目标的图像区域中提取特征信息,并根据所述特征信息构造特征模型;
    确定模块,设置为根据当前帧图像的跟踪结果的可信度确定跟踪目标的跟踪状态,其中,所述跟踪结果的可信度由所述当前帧图像预设区域的特征信息与所述特征模型的相似度确定;
    更新模块,设置为在根据所述跟踪状态确定所述跟踪目标未丢失的情况下,根据所述当前帧图像的跟踪结果更新所述特征模型。
  9. 根据权利要求8所述的装置,其中,所述构造模块包括:
    初始化子模块,设置为通过从所述跟踪目标的图像区域中提取的特征信息替换预设模型中的原始特征信息,得到所述特征模型。
  10. 根据权利要求9所述的装置,其中,所述更新模块包括:
    提取子模块,设置为对获取当前帧图像的跟踪结果的图像区域进行特征提取,并进行归一化处理,得到对应的特征信息;
    第一获取子模块,设置为获取预设概率;
    替换子模块,设置为通过获取的当前帧图像对应的特征信息中的预设特征信 息以预设概率替换所述特征模型中的任意一个特征信息,以更新所述特征模型。
  11. 根据权利要求10所述的装置,其中,所述第一获取子模块包括:
    获取单元,设置为获取所述当前帧图像的特征信息与最近一次更新的特征模型的巴氏距离;
    计算单元,设置为通过如下公式获取所述预设概率:
    Figure PCTCN2017116330-appb-100002
    其中,所述p为所述预设概率,所述dmedian为所述特征信息与所述最近一次更新的特征模型的巴氏距离,所述σ为预设常数。
  12. 根据权利要求11所述的装置,其中,所述获取单元包括:
    确定子单元,设置为确定所述当前帧图像的特征信息与所述最近一次更新的模型中的多个特征信息的多个巴氏距离的中值为所述特征信息与所述最近一次更新的模型的巴氏距离。
  13. 根据权利要求8至12中任意一项所述的装置,其中,所述构造模块包括:
    去除背景子模块,设置为去除所述跟踪目标的图像区域中的背景图像;
    分割子模块,设置为将去除所述背景图像的所述图像区域沿预设方向分割为多个图像;
    第二获取子模块,设置为获取均分后的所述多个图像的特征信息;
    连接子模块,设置为将均分后的所述多个图像的特征信息按照分割的顺序进行连接,得到所述跟踪目标的图像的特征信息。
  14. 根据权利要求8至13中任意一项所述的装置,其中,所述特征信息为图像颜色特征信息,其中,所述图像颜色特征信息包括:颜色名称信息和/色调信息。
  15. 一种存储介质,其特征在于,所述存储介质包括存储的程序,其中,在所述程序运行时控制所述存储介质所在设备执行权利要求1至7中任意一项所述的目标再识别方法。
  16. 一种处理器,其特征在于,所述处理器用于运行程序,其中,所述程序运行时执行权利要求1至7中任意一项所述的目标再识别方法。
PCT/CN2017/116330 2016-12-30 2017-12-15 目标再识别方法和装置 WO2018121287A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201611265859.0A CN108268823B (zh) 2016-12-30 2016-12-30 目标再识别方法和装置
CN201611265859.0 2016-12-30

Publications (1)

Publication Number Publication Date
WO2018121287A1 true WO2018121287A1 (zh) 2018-07-05

Family

ID=62710168

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/116330 WO2018121287A1 (zh) 2016-12-30 2017-12-15 目标再识别方法和装置

Country Status (2)

Country Link
CN (1) CN108268823B (zh)
WO (1) WO2018121287A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109034109A (zh) * 2018-08-16 2018-12-18 新智数字科技有限公司 一种基于聚类算法的行人重识别方法和装置
CN110907909A (zh) * 2019-10-30 2020-03-24 南京市德赛西威汽车电子有限公司 一种基于概率统计的雷达目标识别方法
CN111709978A (zh) * 2020-05-06 2020-09-25 广东康云科技有限公司 一种跨屏目标跟踪方法、***、装置及存储介质
CN111738063A (zh) * 2020-05-08 2020-10-02 华南理工大学 船舶目标跟踪方法、***、计算机设备及存储介质
CN111860190A (zh) * 2020-06-24 2020-10-30 国汽(北京)智能网联汽车研究院有限公司 目标追踪的方法、装置、设备及存储介质
CN112784648A (zh) * 2019-11-07 2021-05-11 中国科学技术大学 一种优化视频行人重识别***特征提取的方法及装置

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111127508B (zh) * 2018-10-31 2023-05-02 杭州海康威视数字技术股份有限公司 一种基于视频的目标跟踪方法及装置
CN109635657B (zh) * 2018-11-12 2023-01-06 平安科技(深圳)有限公司 目标跟踪方法、装置、设备及存储介质
CN111753601B (zh) * 2019-03-29 2024-04-12 华为技术有限公司 一种图像处理的方法、装置以及存储介质
CN113536914A (zh) * 2021-06-09 2021-10-22 重庆中科云从科技有限公司 一种对象跟踪识别方法、***、设备及介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104240266A (zh) * 2014-09-04 2014-12-24 成都理想境界科技有限公司 基于颜色-结构特征的目标对象跟踪方法
CN105335986A (zh) * 2015-09-10 2016-02-17 西安电子科技大学 基于特征匹配和MeanShift算法的目标跟踪方法
CN106023155A (zh) * 2016-05-10 2016-10-12 电子科技大学 基于水平集的在线目标轮廓跟踪方法
CN106056627A (zh) * 2016-05-30 2016-10-26 河海大学 一种基于局部鉴别性稀疏表示的鲁棒目标跟踪方法
CN106097385A (zh) * 2016-05-31 2016-11-09 海信集团有限公司 一种目标跟踪的方法和装置

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6590999B1 (en) * 2000-02-14 2003-07-08 Siemens Corporate Research, Inc. Real-time tracking of non-rigid objects using mean shift
CN101888479B (zh) * 2009-05-14 2012-05-02 汉王科技股份有限公司 检测和跟踪目标图像的方法及装置
CN101901486B (zh) * 2009-11-17 2012-10-03 华为技术有限公司 运动目标检测方法和装置
CN102339390B (zh) * 2011-09-16 2014-06-04 天津市亚安科技股份有限公司 一种视频监控***目标模板的更新方法及***
CN104020466B (zh) * 2014-06-17 2016-05-25 西安电子科技大学 基于变结构多模型的机动目标跟踪方法
CN104036523A (zh) * 2014-06-18 2014-09-10 哈尔滨工程大学 一种基于SURF特征的改进Mean Shift的目标跟踪方法
CN104252178B (zh) * 2014-09-12 2017-11-03 西安电子科技大学 一种基于强机动的目标跟踪方法
CN105654508B (zh) * 2015-12-24 2018-06-01 武汉大学 基于自适应背景分割的监控视频移动目标跟踪方法及***
CN105931269A (zh) * 2016-04-22 2016-09-07 海信集团有限公司 一种视频中的目标跟踪方法及装置
CN105930833B (zh) * 2016-05-19 2019-01-22 重庆邮电大学 一种基于视频监控的车辆跟踪与分割方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104240266A (zh) * 2014-09-04 2014-12-24 成都理想境界科技有限公司 基于颜色-结构特征的目标对象跟踪方法
CN105335986A (zh) * 2015-09-10 2016-02-17 西安电子科技大学 基于特征匹配和MeanShift算法的目标跟踪方法
CN106023155A (zh) * 2016-05-10 2016-10-12 电子科技大学 基于水平集的在线目标轮廓跟踪方法
CN106056627A (zh) * 2016-05-30 2016-10-26 河海大学 一种基于局部鉴别性稀疏表示的鲁棒目标跟踪方法
CN106097385A (zh) * 2016-05-31 2016-11-09 海信集团有限公司 一种目标跟踪的方法和装置

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109034109A (zh) * 2018-08-16 2018-12-18 新智数字科技有限公司 一种基于聚类算法的行人重识别方法和装置
CN109034109B (zh) * 2018-08-16 2021-03-23 新智数字科技有限公司 一种基于聚类算法的行人重识别方法和装置
CN110907909A (zh) * 2019-10-30 2020-03-24 南京市德赛西威汽车电子有限公司 一种基于概率统计的雷达目标识别方法
CN110907909B (zh) * 2019-10-30 2023-09-12 南京市德赛西威汽车电子有限公司 一种基于概率统计的雷达目标识别方法
CN112784648A (zh) * 2019-11-07 2021-05-11 中国科学技术大学 一种优化视频行人重识别***特征提取的方法及装置
CN112784648B (zh) * 2019-11-07 2022-09-06 中国科学技术大学 一种优化视频行人重识别***特征提取的方法及装置
CN111709978A (zh) * 2020-05-06 2020-09-25 广东康云科技有限公司 一种跨屏目标跟踪方法、***、装置及存储介质
CN111738063A (zh) * 2020-05-08 2020-10-02 华南理工大学 船舶目标跟踪方法、***、计算机设备及存储介质
CN111738063B (zh) * 2020-05-08 2023-04-18 华南理工大学 船舶目标跟踪方法、***、计算机设备及存储介质
CN111860190A (zh) * 2020-06-24 2020-10-30 国汽(北京)智能网联汽车研究院有限公司 目标追踪的方法、装置、设备及存储介质
CN111860190B (zh) * 2020-06-24 2024-04-12 国汽(北京)智能网联汽车研究院有限公司 目标追踪的方法、装置、设备及存储介质

Also Published As

Publication number Publication date
CN108268823B (zh) 2021-07-20
CN108268823A (zh) 2018-07-10

Similar Documents

Publication Publication Date Title
WO2018121287A1 (zh) 目标再识别方法和装置
CN109145766B (zh) 模型训练方法、装置、识别方法、电子设备及存储介质
CN109977262B (zh) 从视频中获取候选片段的方法、装置及处理设备
CN107346409B (zh) 行人再识别方法和装置
WO2018121286A1 (zh) 目标跟踪方法和装置
CN108229322B (zh) 基于视频的人脸识别方法、装置、电子设备及存储介质
Noh et al. A new framework for background subtraction using multiple cues
US9129191B2 (en) Semantic object selection
CN108288051B (zh) 行人再识别模型训练方法及装置、电子设备和存储介质
US9129192B2 (en) Semantic object proposal generation and validation
US20160307057A1 (en) Fully Automatic Tattoo Image Processing And Retrieval
WO2019128254A1 (zh) 图像分析方法、装置、电子设备及可读存储介质
WO2017181892A1 (zh) 前景分割方法及装置
CN110019895B (zh) 一种图像检索方法、装置及电子设备
CN112016531A (zh) 模型训练方法、对象识别方法、装置、设备及存储介质
WO2016139964A1 (ja) 注目領域抽出装置および注目領域抽出方法
WO2019197021A1 (en) Device and method for instance-level segmentation of an image
CN106127222B (zh) 一种基于视觉的字符串相似度计算方法及相似性判断方法
US10110846B2 (en) Computationally efficient frame rate conversion system
Banerjee et al. Iris segmentation using geodesic active contours and grabcut
CN112633221A (zh) 一种人脸方向的检测方法及相关装置
CN112036362A (zh) 图像处理方法、装置、计算机设备和可读存储介质
CN112270204A (zh) 目标识别方法及装置、存储介质及电子设备
KR102166117B1 (ko) 시멘틱 매칭 장치 및 방법
CN111079757A (zh) 服饰属性识别方法、装置及电子设备

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: 17887160

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: 17887160

Country of ref document: EP

Kind code of ref document: A1