CN115731382A - Point cloud target detection method and device, computer equipment and storage medium - Google Patents
Point cloud target detection method and device, computer equipment and storage medium Download PDFInfo
- Publication number
- CN115731382A CN115731382A CN202211439886.0A CN202211439886A CN115731382A CN 115731382 A CN115731382 A CN 115731382A CN 202211439886 A CN202211439886 A CN 202211439886A CN 115731382 A CN115731382 A CN 115731382A
- Authority
- CN
- China
- Prior art keywords
- grid
- target
- feature
- point cloud
- detection
- Prior art date
- Legal status (The legal status 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 status listed.)
- Pending
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The application relates to a point cloud target detection method, a point cloud target detection device, a computer device and a storage medium. The method comprises the following steps: the method comprises the steps of obtaining point cloud data, carrying out rasterization processing on the point cloud data, carrying out dimension raising and aggregation on features of points in each grid to obtain corresponding grid features, generating feature maps according to the grid features of each grid, carrying out feature weighting on the feature maps and carrying out target detection to obtain target information. By adopting the method, the grid characteristics can be fully extracted, the characteristics in each grid can obtain global attention, the detection accuracy of the target is effectively improved, and the high-efficiency perception of the environment is realized.
Description
Technical Field
The present application relates to the field of radar technologies, and in particular, to a method and an apparatus for detecting a point cloud target, a computer device, and a storage medium.
Background
At present, laser radars are widely concerned by researchers due to the characteristics of high detection precision, strong interference resistance and the like, and gradually become one of unmanned core sensors. The point cloud is a representation form of data acquired by the laser radar, is a set formed by points which are distributed in a target or scene space and are expressed in the same coordinate system, and can effectively reflect three-dimensional information of a real scene. However, the point cloud has the characteristics of disorder and unstructured characteristics, and also brings problems to data processing.
The existing laser radar target detection methods can be divided into two types: one is a target detection method based on a traditional algorithm, and the other is a target detection method based on deep learning. The target detection method based on the traditional algorithm is mature, but the intelligent level is low, the extracted semantic category is single, and when the target detection method faces a complex street scene, error recognition is easy to occur due to special obstacles such as a wall surface and a road edge, and the target detection method based on the deep learning does not need a large number of points in the target detection process.
Disclosure of Invention
In view of the foregoing, it is necessary to provide a point cloud object detection method, apparatus, computer device and storage medium capable of improving object detection accuracy.
In a first aspect, the present application provides a point cloud target detection method, including:
acquiring point cloud data;
rasterizing the point cloud data;
performing dimension increasing and aggregation on the features of the points in each grid to obtain corresponding grid features, and generating a feature map according to the grid features of each grid;
and carrying out feature weighting on the feature map and carrying out target detection to obtain target information.
In one embodiment, rasterizing the point cloud data includes:
and performing grid division and screening on the point cloud data according to a preset grid threshold value.
In one embodiment, performing dimension-raising and aggregation on the features of the points in each grid to obtain the corresponding grid features includes:
performing dimensionality increase on the characteristics of each grid inner point to obtain high-dimensional characteristic information, and aggregating the characteristics of each grid inner point through maximum pooling to obtain a first-layer characteristic;
and splicing the high-dimensional characteristic information and the first layer characteristics, and performing dimensionality increase and aggregation to obtain second layer characteristics of each grid, wherein the second layer characteristics are used as grid characteristics.
In one embodiment, the performing feature weighting and target detection on the feature map and obtaining target information includes:
performing convolution on the feature graph to obtain feature subgraphs of different sizes;
expanding each characteristic subgraph to obtain a corresponding one-dimensional characteristic vector;
weighting each one-dimensional feature vector and restoring the one-dimensional feature vector to a feature subgraph;
and splicing the weighted feature sub-graphs, and performing target detection on the spliced feature graphs to obtain target information.
In one embodiment, in the training process of the neural network model for target detection, the detection frame with the maximum target class value is used as the position result of the target information, the class with the maximum target class value is used as the target class of the target information, the target class value is used as the confidence coefficient of the detection frame of the target information, the output target information is compared with the actual result, the loss is returned, and the parameters in the whole model are updated.
In one embodiment, after performing feature weighting on the feature map and performing target detection, the method further includes:
and screening the target information through non-maximum value inhibition to obtain a final target.
In one embodiment, screening the target information through non-maximum suppression to obtain a final target includes:
screening target information according to a target category value in the target information and a preset first threshold value to obtain an effective detection frame;
arranging the effective detection frames according to the confidence degree of the detection frames in the target information, and calculating the overlap-intersection parallel ratio of the front effective detection frame and the rear effective detection frame;
and screening the effective detection frames according to the overlapping intersection ratio and a preset second threshold value to obtain a final target.
In a second aspect, the present application further provides a point cloud target detection apparatus, including:
the acquisition module is used for acquiring point cloud data;
the rasterization module is used for rasterizing the point cloud data;
the characteristic extraction module is used for performing dimension increasing and aggregation on the characteristics of the inner points of each grid through the multilayer perceptron to obtain corresponding grid characteristics, and generating a characteristic diagram according to the grid characteristics of each grid;
and the target detection module is used for carrying out characteristic weighting on the characteristic diagram and carrying out target detection to obtain target information.
In a third aspect, the present application further provides a computer device, including a memory and a processor, where the memory stores a computer program, and the processor implements the steps of the point cloud target detection method described in any one of the above embodiments when executing the computer program.
In a fourth aspect, the present application further provides a computer-readable storage medium, on which a computer program is stored, where the computer program is executed by a processor to implement the steps of the point cloud target detection method according to any one of the above embodiments.
According to the point cloud target detection method, the point cloud data are obtained, rasterization processing is carried out on the point cloud data, dimension increasing and aggregation are carried out on the features of points in each grid to obtain the corresponding grid features, the feature graph is generated according to the grid features of each grid, feature weighting is carried out on the feature graph, and target detection is carried out to obtain target information.
Drawings
In order to more clearly illustrate the technical solutions in the embodiments or the conventional technologies of the present application, the drawings used in the descriptions of the embodiments or the conventional technologies will be briefly introduced below, it is obvious that the drawings in the following descriptions are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a schematic flow chart of a method for detecting a point cloud object according to an embodiment;
FIG. 2 is a schematic diagram illustrating the processing of step S200 in the point cloud target detection method according to an embodiment;
FIG. 3 is a schematic flow chart illustrating step S300 of the point cloud target detection method according to an embodiment;
FIG. 4 is a logic diagram illustrating the step S300 of the point cloud object detection method according to an embodiment
FIG. 5 is a schematic flow chart illustrating step S400 of the point cloud target detection method according to an embodiment;
FIG. 6 is a logic diagram illustrating step S400 of the point cloud object detection method in one embodiment;
FIGS. 7 to 9 are schematic diagrams illustrating test results of the point cloud object detection method according to an embodiment;
FIG. 10 is a block diagram of an embodiment of a point cloud target detection apparatus;
FIG. 11 is a diagram of the internal structure of a computer device in one embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the description of the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.
It will be understood that, as used herein, the terms "first," "second," and the like may be used herein to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish one element from another.
It will be understood that when an element is referred to as being "connected" to another element, it can be directly connected to the other element or be connected to the other element through intervening elements. In addition, "connection" in the following embodiments is understood to mean "electrical connection", "communication connection", and the like if there is a transfer of electrical signals or data between the connected objects.
As used herein, the singular forms "a", "an" and "the" may include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises/comprising," "includes" or "including," or "having," and the like, specify the presence of stated features, integers, steps, operations, components, parts, or combinations thereof, but do not preclude the presence or addition of one or more other features, integers, steps, operations, components, parts, or combinations thereof.
The point cloud target detection method provided by the embodiment of the application can be applied to laser radars to realize target detection, and especially can be used for target detection in the field of unmanned driving. The point cloud is a representation form of data acquired by the laser radar, is a set formed by points which are distributed in a target or scene space and are expressed in the same coordinate system, and can effectively reflect three-dimensional information of a real scene. However, a point cloud has the characteristics of disorder and unstructured, wherein the order of the disordered points does not affect the representation of the distribution of the points in the whole body shape or scene in space, the positions of the points in space are fixed, but the points can be arranged in different orders, so that the same point cloud graph can have a plurality of completely different matrixes, and unstructured means that the point cloud is composed of the points distributed in a three-dimensional space, but the connection information between the points is not clear, so that the uncertainty of the point cloud structure is caused.
Therefore, a point cloud target detection solution is provided to improve the accuracy of laser radar target detection and realize efficient environmental perception.
In one embodiment, as shown in fig. 1, a method for detecting a point cloud target is provided, which includes the following steps:
s100: acquiring point cloud data;
specifically, the embodiment scans through the laser radar to obtain the point cloud data, wherein the point cloud data takes the three-dimensional coordinates and the reflectivity of the point cloud as the main data for detecting the laser radar target.
Optionally, the point cloud data are filtered according to a preset area range, point clouds far away from the radar are filtered, and laser point cloud data in a specified range are reserved.
S200: rasterizing the point cloud data;
specifically, if a scene is subjected to three-dimensional voxel division, a large amount of blank spaces without points exist in the spaces, and the calculation amount in the later period is increased, for this reason, referring to fig. 2, in this embodiment, point cloud data is subjected to rasterization processing, point cloud is subjected to columnar division into grids, the grids are screened, a required grid is reserved, and feature extraction is further performed based on the reserved grid internal points.
Optionally, the grid is screened according to a preset grid threshold, if the number of points in the grid is greater than the grid threshold, the grid is retained, and the position of the grid is recorded.
Alternatively, [ x, y, z, r, x _ c, y _ c, z _ c, x _ p, y _ p ] can be used to represent the characteristics of points within the grid, x, y, z, r is the 4-dimensional data of the original point, x _ c, y _ c, z _ c represent the deviation of the grid midpoint from the grid centroid, x _ p, y _ p refer to the deviation values from the grid center in the xy plane.
S300: performing dimension raising and aggregation on the features of the points in each grid to obtain corresponding grid features, and generating a feature map according to the grid features of each grid;
specifically, in order to improve the effect of extracting the grid features, in the embodiment, a transform attention mechanism is added to the grid feature extraction, the features of points in the grid are aggregated through a transform module, and a multi-layer Perceptron (MLP) is combined to perform dimension enhancement to express the grid features, so that the feature extraction in the grid is more sufficient, wherein the transform module is originally applied to natural language processing to capture long-term information and dependency relationships between sequence elements, the correlation between every two entities is predicted through the form of the self-attention mechanism, and the multi-layer Perceptron can be understood as a one-dimensional neural network, and dimension enhancement and dimension reduction of vectors can be realized through a plurality of sets of weights, and the multi-layer Perceptron is generally used for expansion and compression of feature vectors.
Specifically, the extracted grid features are rearranged on a two-dimensional plane according to the positions of the grids to obtain a required feature map, for example, the feature map with the dimension of 512 × 64, the dimension of 512 × 512 is the size of the xy plane, and the dimension of 64 is the feature dimension, so that the conversion from the three-dimensional point cloud to the two-dimensional feature map is completed, and the point cloud is regularized.
S400: and carrying out feature weighting and target detection on the feature map to obtain target information.
Specifically, in order to improve the accuracy of target detection, in the embodiment, feature weighting is performed on the feature map in target detection, wherein the feature map is specifically operated through a transform module, it is worth noting that features are not aggregated in the process, but the weighted features are restored back to the original feature size, so that the original feature map can obtain global association according to the position information, that is, the features in each grid obtain global attention, and more reliable features are provided for subsequent target detection.
Specifically, RPN (Region Proposal Network) Network detection is further performed based on the weighted feature map, and target information is obtained through output, wherein the RPN Network is used for extracting detection frames, the grid features on each two-dimensional plane are detected through the RPN Network, the features are converted into a plurality of target information through a neural Network, and the target information comprises a target category, three-dimensional position coordinates (x, y, z) of the detection frames, the sizes, the widths, the heights (l, w, h) of the detection frames and orientation information r of the detection frames. In addition to the RPN network, other neural networks may be used for target detection in this embodiment, such as R-CNN, fast R-CNN, etc.
In the point cloud target detection method, on one hand, dimension increasing and aggregation are performed in grid feature extraction, so that feature extraction in grids is more sufficient, and the detection accuracy of a target is improved, on the other hand, feature weighting is performed on feature maps in target detection, so that features in each grid obtain global attention, more reliable features are provided for subsequent target detection, the detection accuracy of the target is further improved, and efficient environment perception is achieved.
In one embodiment, referring to fig. 3 and 4, the upscaling and aggregating the features of the points in each grid to obtain the corresponding grid features comprises: s301: performing dimensionality increase on the characteristics of each grid inner point to obtain high-dimensional characteristic information, and aggregating the characteristics of each grid inner point through maximum pooling to obtain a first-layer characteristic; s302: and splicing the high-dimensional feature information and the first layer features, performing dimension increasing and aggregation to obtain second layer features of each grid, and taking the second layer features as the grid features.
Specifically, referring to fig. 4, the present embodiment adds a transform attention mechanism to the grid feature extraction, specifically:
firstly, the coordinates and reflectivity characteristics of each grid inner point are subjected to dimension increase through a multilayer perceptron to obtain high-dimensional characteristic information, and the characteristics of each grid inner point are aggregated according to the maximum dimension through maximum value pooling to form a first-layer characteristic;
and secondly, splicing the first layer of features with the original high-dimensional feature information, inputting the high-dimensional feature information into the multilayer perceptron again for dimensionality increase, adding the spliced and dimensionality-increased feature information into a transducer module, and aggregating the point features again to obtain a second layer of features, namely grid features.
Specifically, the main calculation formula of Transform is:
wherein Q represents a query correlation matrix, K represents a key weight matrix, V represents a value matrix, and d represents the dimensions of Q and K, so as to balance the fact that the variance is larger and larger as the dimensions of Q and K are increased.
And obtaining weight incidence matrixes of all points by the inner product of Q and K, then carrying out normalization operation through softmax, namely obtaining attention weight, then multiplying the weight incidence matrixes by the V matrix to obtain weight of the V matrix, finally adding the weight incidence matrixes according to dimensionality to obtain perfect second-layer characteristics, and taking the second-layer characteristics as grid characteristics of subsequent operation.
In one embodiment, referring to fig. 5, performing feature weighting on the feature map and performing target detection, and obtaining target information includes: s401: performing convolution on the feature graph to obtain feature subgraphs of different sizes; s402: expanding each characteristic subgraph to obtain a corresponding one-dimensional characteristic vector; s403: weighting each one-dimensional feature vector and restoring the one-dimensional feature vector to a feature subgraph; s404: and splicing the weighted characteristic subgraphs, and carrying out target detection on the spliced characteristic subgraphs to obtain target information.
Specifically, the feature graph is converted into feature sub-graphs of different sizes by performing convolution operation on the feature graph, each feature sub-graph is expanded into a one-dimensional feature vector, then each one-dimensional feature vector of the feature graph is weighted again through a Transformer and restored to the original feature size, so that weighting of the feature graph is completed, and the original feature graph can obtain global correlation according to position information.
For example, a one-dimensional feature matrix is extracted from the feature map according to the z-axis dimension, the feature matrix is expanded according to the rows to form a one-dimensional feature vector, Q, K and V matrices are respectively trained through a multilayer perceptron, then the weighted one-dimensional feature vector is obtained based on a Transform main calculation formula, the features are restored back to the original feature matrix according to the positions of the rows, and the weighting of the feature matrix is completed once. The above operation is performed for each dimension of the feature map in the z-axis, so that the feature map is weighted.
Specifically, after weighting the feature subgraphs of different sizes, splicing the features along the z-axis direction to enable the features to have more comprehensive information, and finally, performing RPN network detection on the spliced feature graphs and outputting to obtain target information.
Specifically, the grid feature outputs a plurality of sets of target information through the RPN network, each set of target information including a target category, three-dimensional position coordinates (x, y, z), a size, a length, a width and a height (l, w, h) of a detection frame, and detection frame orientation information r.
In one embodiment, in the training process of the neural network model for target detection, the detection frame with the maximum target class value is used as the position result of the target information, the class with the maximum target class value is used as the target class of the target information, the target class value is used as the confidence coefficient of the detection frame of the target information, the output target information is compared with the actual result, the loss is returned, and the parameters in the whole model are updated.
The loss function is as follows:
Δθ＝sin(θ gt -θ a )
loss cls ＝-α(1-p) γ log p
loss dir ＝∑-(dir gt log dir a +(1-dir gt )log(1-dir a ))
wherein, beta loc ＝2，β cls ＝1，β dir Where "= 0.2, Δ x, Δ y, Δ z, Δ h, Δ w, Δ l, Δ θ represent the amount of deviation of coordinates, dimensions, and angles, respectively, and since Δ θ is calculated to be 0 if two orientation angles differ by an integral multiple of 180 degrees, the orientation angles are divided into four imagings [0, 90 ]]，[90，180]，[180，270]，[270，360]，loss dir The method is a cross entropy function to calculate the loss of the orientation angle in the classification process, so that the orientation of the angle is more accurate. And finally, completing the whole parameter training of the model by using a gradient descent mode.
In one embodiment, referring to fig. 1, after performing feature weighting on the feature map and performing target detection, the method further includes: s500: and screening the target information through non-maximum value inhibition to obtain a final target.
Specifically, the obtained target information includes a great amount of detection frame information, and many of the detection frame information are background types and are not used, so that the present embodiment screens the detection frames by non-maximum suppression: firstly, screening a threshold value of a category value, and taking the position target as an effective detection frame when the category value of the position target is greater than a certain threshold value; then arranging the detection frames according to the confidence level, firstly queuing the detection frame with the highest confidence level, overlapping and comparing the detection frame with the subsequent detection frame, inputting the queue if the overlapping and comparing ratio is less than a certain threshold value, and repeating the operation; and finally, outputting the detection frame in the queue, namely, taking the detection frame as a final target detection output result.
Referring to fig. 7 to 9, in this embodiment, for the relationship between the P-R curve graph, the P-R curve graph reaction model detection accuracy and the recall ratio of the car, the cycle and the pedestrian detection, respectively, from the viewpoint of the test result, under different detection difficulties, the point cloud target detection method based on the above embodiments all obtains good performance, and completes the self-attention performance of the point cloud.
In addition, in this embodiment, the traditional target detection method pointpilars based on deep learning and the point cloud target detection method in this embodiment are compared and tested on kitti, the comparison index refers to AP of kitti, and the test results under the same experimental conditions are as follows:
| PointPillar(％) | Ours(％) | |
| Car | 86.9405 | 88.9718 |
| Pedestrian | 52.1356 | 55.2964 |
| Cyclist | 79.9051 | 82.3727 |
the KITTI data set is a data set of the laser radar, the AP (access precision) is used for reflecting the detection capability of the model for different targets, and from the test result, on the basis of the detection capability of three different targets of car, cyclist and pedestrian, the point cloud target detection method in the embodiment obtains better performance and completes the self-attention performance of the point cloud.
It should be understood that, although the steps in the flowcharts related to the embodiments are shown in sequence as indicated by the arrows, the steps are not necessarily executed in sequence as indicated by the arrows. The steps are not performed in the exact order shown and described, and may be performed in other orders, unless explicitly stated otherwise. Moreover, at least a part of the steps in the flowcharts related to the above embodiments may include multiple steps or multiple stages, which are not necessarily performed at the same time, but may be performed at different times, and the order of performing the steps or stages is not necessarily sequential, but may be performed alternately or alternately with other steps or at least a part of the steps or stages in other steps.
Based on the same inventive concept, the embodiment of the application also provides a point cloud target detection device for realizing the point cloud target detection method. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the method, so the specific limitations in one or more embodiments of the point cloud target detection device provided below can refer to the limitations on the point cloud target detection method in the above, and details are not repeated here.
In one embodiment, as shown in fig. 10, there is provided a point cloud target detection apparatus including:
the acquisition module 10 is used for acquiring point cloud data;
the rasterizing module 20 is configured to perform rasterization processing on the point cloud data;
the feature extraction module 30 is configured to perform dimension raising and aggregation on features of points in each grid through a multilayer perceptron to obtain corresponding grid features, and generate a feature map according to the grid features of each grid;
and the target detection module 40 is used for performing feature weighting on the feature map and performing target detection to obtain target information.
In one embodiment, rasterizing the point cloud data includes: and performing grid division and screening on the point cloud data according to a preset grid threshold value.
In one embodiment, the step of performing dimension-raising and aggregation on the features of the points in each grid to obtain the corresponding grid features includes: performing dimensionality increase on the characteristics of each grid inner point to obtain high-dimensional characteristic information, and aggregating the characteristics of each grid inner point through maximum pooling to obtain a first-layer characteristic; and splicing the high-dimensional characteristic information and the first layer characteristics, and performing dimensionality increase and aggregation to obtain second layer characteristics of each grid, wherein the second layer characteristics are used as grid characteristics.
In one embodiment, the performing feature weighting and target detection on the feature map and obtaining target information includes: performing convolution on the feature graph to obtain feature subgraphs of different sizes; expanding each characteristic subgraph to obtain a corresponding one-dimensional characteristic vector; weighting each one-dimensional feature vector and restoring the one-dimensional feature vector to a feature subgraph; and splicing the weighted characteristic subgraphs, and carrying out target detection on the spliced characteristic subgraphs to obtain target information.
In one embodiment, in the training process of the neural network model for target detection, the detection frame with the maximum target class value is used as the position result of the target information, the class with the maximum target class value is used as the target class of the target information, the target class value is used as the confidence coefficient of the detection frame of the target information, the output target information is compared with the actual result, the loss is returned, and the parameters in the whole model are updated.
In one embodiment, referring to fig. 10, after the feature weighting and the target detection are performed on the feature map, the method further includes: and the output module 50 is used for screening the target information through non-maximum suppression to obtain a final target.
In one embodiment, screening the target information by non-maxima suppression to obtain a final target comprises: screening target information according to a target category value in the target information and a preset first threshold value to obtain an effective detection frame; arranging the effective detection frames according to the confidence degree of the detection frames in the target information, and calculating the overlapping and intersection ratio of the front effective detection frame and the rear effective detection frame; and screening effective detection frames according to the overlapping intersection ratio and a preset second threshold value to obtain a final target.
All or part of the modules in the point cloud target detection device can be realized by software, hardware and a combination thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.
In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as shown in fig. 11. The computer device comprises a processor, a memory, a communication interface, a display screen and an input device which are connected through a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a nonvolatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program to run on the non-volatile storage medium. The communication interface of the computer device is used for carrying out wired or wireless communication with an external terminal, and the wireless communication can be realized through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. The computer program is executed by a processor to implement a point cloud object detection method. The display screen of the computer equipment can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer equipment can be a touch layer covered on the display screen, a key, a track ball or a touch pad arranged on the shell of the computer equipment, an external keyboard, a touch pad or a mouse and the like.
Those skilled in the art will appreciate that the architecture shown in fig. 11 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.
In one embodiment, a computer device is provided, which includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement any one of the point cloud target detection methods in the above embodiments. For a detailed description, reference is made to the corresponding description of the method, which is not repeated herein.
In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which when executed by a processor implements any one of the point cloud target detection methods in the above embodiments. The detailed description refers to the corresponding description of the method, and is not repeated herein.
It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above may be implemented by hardware instructions of a computer program, which may be stored in a non-volatile computer-readable storage medium, and when executed, may include the processes of the embodiments of the methods described above. Any reference to memory, database, or other medium used in the embodiments provided herein may include at least one of non-volatile and volatile memory. The nonvolatile Memory may include Read-Only Memory (ROM), magnetic tape, floppy disk, flash Memory, optical Memory, high-density embedded nonvolatile Memory, resistive Random Access Memory (ReRAM), magnetic Random Access Memory (MRAM), ferroelectric Random Access Memory (FRAM), phase Change Memory (PCM), graphene Memory, and the like. Volatile Memory can include Random Access Memory (RAM), external cache Memory, and the like. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM), among others.
The processors referred to in the embodiments provided herein may be general purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, quantum computing based data processing logic devices, etc., without limitation.
All possible combinations of the technical features in the above embodiments may not be described for the sake of brevity, but should be considered as being within the scope of the present disclosure as long as there is no contradiction between the combinations of the technical features.
The above examples only express several embodiments of the present application, and the description thereof is more specific and detailed, but not to be construed as limiting the scope of the present application. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, and these are all within the scope of protection of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.
Claims (10)
1. A point cloud target detection method, the method comprising:
acquiring point cloud data;
rasterizing the point cloud data;
performing dimension increasing and aggregation on the features of the points in each grid to obtain corresponding grid features, and generating a feature map according to the grid features of each grid;
and carrying out feature weighting on the feature graph and carrying out target detection to obtain target information.
2. The method of claim 1, wherein rasterizing the point cloud data comprises:
and performing grid division and screening on the point cloud data according to a preset grid threshold value.
3. The method of claim 1, wherein the step of performing dimension-raising and aggregation on the features of the points in each grid to obtain the corresponding grid features comprises:
performing dimensionality increase on the characteristics of each grid inner point to obtain high-dimensional characteristic information, and aggregating the characteristics of each grid inner point through maximum pooling to obtain a first-layer characteristic;
and splicing the high-dimensional feature information and the first layer features, performing dimension increasing and aggregation to obtain second layer features of each grid, and taking the second layer features as the grid features.
4. The method of claim 1, wherein the performing feature weighting and object detection on the feature map to obtain object information comprises:
performing convolution on the feature graph to obtain feature subgraphs of different sizes;
expanding each characteristic subgraph to obtain a corresponding one-dimensional characteristic vector;
weighting each one-dimensional feature vector and restoring the one-dimensional feature vector to the feature subgraph;
and splicing the weighted feature subgraphs, and carrying out target detection on the spliced feature subgraphs to obtain the target information.
5. The method according to claim 4, wherein in the training process of the neural network model for target detection, the detection frame with the maximum target class value is used as the position result of the target information, the class with the maximum target class value is used as the target class of the target information, the target class value is used as the confidence of the detection frame of the target information, the output target information is compared with the actual result, and the loss is returned to update the parameters in the whole model.
6. The method according to any one of claims 1 to 5, wherein after the feature weighting and the target detection are performed on the feature map, the method further comprises:
and screening the target information through non-maximum value inhibition to obtain a final target.
7. The method of claim 6, wherein the screening the target information for a final target by non-maxima suppression comprises:
screening the target information according to a target category value in the target information and a preset first threshold value to obtain an effective detection frame;
arranging the effective detection frames according to the confidence degree of the detection frames in the target information, and calculating the overlapping and comparing ratio of the effective detection frames before and after the overlapping and comparing;
and screening the effective detection frame according to the overlapping intersection ratio and a preset second threshold value to obtain the final target.
8. A point cloud target detection apparatus, the apparatus comprising:
the acquisition module is used for acquiring point cloud data;
the rasterization module is used for rasterizing the point cloud data;
the characteristic extraction module is used for performing dimension increasing and aggregation on the characteristics of the inner points of each grid through the multilayer perceptron to obtain corresponding grid characteristics, and generating a characteristic diagram according to the grid characteristics of each grid;
and the target detection module is used for carrying out feature weighting on the feature map and carrying out target detection to obtain target information.
9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor, when executing the computer program, implements the steps of the method of any one of claims 1 to 7.
10. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the method of any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211439886.0A CN115731382A (en) | 2022-11-17 | 2022-11-17 | Point cloud target detection method and device, computer equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211439886.0A CN115731382A (en) | 2022-11-17 | 2022-11-17 | Point cloud target detection method and device, computer equipment and storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115731382A true CN115731382A (en) | 2023-03-03 |
Family
ID=85296202
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211439886.0A Pending CN115731382A (en) | 2022-11-17 | 2022-11-17 | Point cloud target detection method and device, computer equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115731382A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117576150A (en) * | 2023-11-03 | 2024-02-20 | 扬州万方科技股份有限公司 | Multi-mode multi-target 3D tracking method and device considering far-frame dependency relationship |
| CN117612070A (en) * | 2024-01-19 | 2024-02-27 | 福思(杭州)智能科技有限公司 | Static truth value data correction method and device and storage medium |
-
2022
- 2022-11-17 CN CN202211439886.0A patent/CN115731382A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117576150A (en) * | 2023-11-03 | 2024-02-20 | 扬州万方科技股份有限公司 | Multi-mode multi-target 3D tracking method and device considering far-frame dependency relationship |
| CN117612070A (en) * | 2024-01-19 | 2024-02-27 | 福思(杭州)智能科技有限公司 | Static truth value data correction method and device and storage medium |
| CN117612070B (en) * | 2024-01-19 | 2024-05-03 | 福思(杭州)智能科技有限公司 | Static truth value data correction method and device and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110059554B (en) | Multi-branch target detection method based on traffic scene | |
| CN115731382A (en) | Point cloud target detection method and device, computer equipment and storage medium | |
| CN110991311B (en) | Target detection method based on dense connection deep network | |
| CN109932730B (en) | Laser radar target detection method based on multi-scale monopole three-dimensional detection network | |
| CN111047078B (en) | Traffic characteristic prediction method, system and storage medium | |
| CN111028327A (en) | Three-dimensional point cloud processing method, device and equipment | |
| EP3965071A2 (en) | Method and apparatus for pose identification | |
| US20210089923A1 (en) | Icospherical gauge convolutional neural network | |
| CN112184867A (en) | Point cloud feature extraction method, device, equipment and storage medium | |
| CN113627440A (en) | Large-scale point cloud semantic segmentation method based on lightweight neural network | |
| CN116246119A (en) | 3D target detection method, electronic device and storage medium | |
| CN115995042A (en) | Video SAR moving target detection method and device | |
| CN111765883B (en) | Robot Monte Carlo positioning method, equipment and storage medium | |
| CN113160117A (en) | Three-dimensional point cloud target detection method under automatic driving scene | |
| Li et al. | Nonlinear von Mises–Fisher filtering based on isotropic deterministic sampling | |
| WO2022017129A1 (en) | Target object detection method and apparatus, electronic device, and storage medium | |
| CN116152800A (en) | 3D dynamic multi-target detection method, system and storage medium based on cross-view feature fusion | |
| CN116129234A (en) | Attention-based 4D millimeter wave radar and vision fusion method | |
| CN110516761A (en) | Object detection system, method, storage medium and terminal based on deep learning | |
| CN115908318A (en) | Power line sag determination method, device, equipment, medium and product | |
| Wang et al. | FPA-DNN: a forward propagation acceleration based deep neural network for ship detection | |
| CN114821033A (en) | Three-dimensional information enhanced detection and identification method and device based on laser point cloud | |
| CN112287995B (en) | Low-resolution image recognition method based on multi-layer coupling mapping | |
| CN113903016A (en) | Bifurcation point detection method, bifurcation point detection device, computer equipment and computer program product | |
| CN115544191A (en) | Three-dimensional point cloud crowdsourcing type semantic map updating method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |