Target detection method, storage medium, electronic device, and vehicle

Abstract

The present application belongs to the field of electric vehicles, and specifically provides a target detection method, a storage medium, an electronic device and a vehicle, aiming to solve the problem of how to improve the speed, generalization and anti-interference ability of point cloud data processing, and realize fast and accurate three-dimensional target detection. For this purpose, the method of the present application comprises: voxelizing the point cloud data to be processed, and obtaining a target detection result through a target detection network composed of a sparse 3D convolution network, a two-dimensional detection network and a detection head network in turn. By voxelizing the point cloud data based on the voxel position, the dependence of the point cloud data on the absolute distance is reduced, the anti-interference ability and the generalization of the network are improved; by optimizing the structure under the sparse 3D convolution network, the detection range in the height direction is expanded without increasing the calculation amount; by introducing the first orientation attribute and the second orientation attribute of the heading angle, the heading angle generation method is improved, and the accuracy of the heading angle detection is improved.

Description

Technical Field

[0001] This invention belongs to the field of electric vehicles, and specifically provides a target detection method, a storage medium, an electronic device, and a vehicle. Background Technology

[0002] With the rapid development of computing, sensing, and networking technologies, many automakers are exploring driver assistance features in vehicles. These features aim to reduce driver intervention, making driving easier and safer, and allowing drivers to fully experience the convenience brought by technological advancements. In driver assistance applications, vehicles need to constantly perceive their surroundings, collect data, and detect static and dynamic objects. Due to the physical characteristics of radar, it can effectively capture shape and distance information of three-dimensional targets; therefore, radar-based target detection methods are receiving increasing attention. Simultaneously, radar-based target detection technology is also being used in applications such as autonomous robot navigation and autonomous vehicle driving.

[0003] Point cloud data acquired by radar is characterized by its massive volume, uneven spatial distribution, and large amount of interference information. Therefore, improving the speed, generalization, and anti-interference ability of point cloud data processing to achieve fast and accurate 3D target detection has become an urgent problem to be solved.

[0004] Accordingly, a new solution is needed in this field to address the aforementioned problems. Summary of the Invention

[0005] The present invention aims to solve or partially solve the above-mentioned technical problems, namely, how to improve the processing speed, generalization and anti-interference ability of point cloud data, and achieve fast and accurate three-dimensional target detection.

[0006] In a first aspect, the present invention provides a target detection method, characterized in that the method comprises:

[0007] Acquire point cloud data;

[0008] The point cloud data is voxelized to obtain voxelized point cloud data;

[0009] Based on the voxelized point cloud data, a 3D point cloud feature map is obtained through a 3D feature extraction network.

[0010] Based on the three-dimensional point cloud feature map, a two-dimensional point cloud feature map is obtained through a two-dimensional detection network;

[0011] Based on the two-dimensional point cloud feature map, the target category and a three-dimensional detection box containing the heading angle are obtained through the detection head network, wherein the heading angle is generated based on the heading angle value and the heading angle orientation attribute.

[0012] In one embodiment of the above target detection method, "obtaining the target category and a 3D detection box including the heading angle through the detection head network based on the two-dimensional point cloud feature map" includes:

[0013] The target category, the three-dimensional detection box, the heading angle value, and the heading angle orientation feature are obtained from the two-dimensional point cloud feature map.

[0014] The heading angle orientation attribute is obtained based on the heading angle orientation feature, and the heading angle orientation attribute includes a first orientation attribute and a second orientation attribute;

[0015] The heading angle is obtained based on the heading angle value, the first heading attribute, and the second heading attribute.

[0016] In one embodiment of the above target detection method, the first orientation attribute includes forward and backward directions, and the second orientation attribute includes left and right directions.

[0017] "Obtaining the heading angle based on the heading angle value, the first heading attribute, and the second heading attribute" includes:

[0018] When the heading angle value is within the first angle range, the heading angle is obtained according to the first orientation attribute;

[0019] When the heading angle value is within the second angle range, the heading angle is obtained according to the second orientation attribute;

[0020] When the heading angle value is within the third angle range, the heading angle is obtained according to the second orientation attribute;

[0021] When the heading angle value is the first boundary angle or the second boundary angle, the heading angle is the position corresponding to the heading angle value;

[0022] The heading angle values ​​are taken in a Cartesian coordinate system (x, y), with the heading angle value in the positive x-axis direction being 0, clockwise rotation being negative, and counterclockwise rotation being positive. The first angle range is [-45°, 45°], the second angle range is (45°, 90°), and the third angle range is (-90°, -45°). The first boundary angle is -90°, and the second boundary angle is 90°.

[0023] In one embodiment of the above target detection method, "voxarizing the point cloud data to obtain voxelized point cloud data" includes:

[0024] Obtain the mean coordinates of each dimension of all point cloud points within each voxel;

[0025] Subtracting the coordinates of the corresponding dimension of the center position of the voxel from the mean of the coordinates yields the voxelized point cloud dimensional features.

[0026] The difference between the mean coordinate of each dimension and the minimum radar detection range corresponding to that dimension is obtained. The difference is then divided by the voxel size of the corresponding dimension to obtain the voxelized point cloud coordinates.

[0027] In one embodiment of the above target detection method, the three-dimensional feature extraction network is a sparse 3D convolutional network, which uses convolutional layers with a kernel size of 3, a stride of 3, and padding of 0 in the height dimension for downsampling.

[0028] In one embodiment of the above target detection method, the method further includes:

[0029] The target detection results are analyzed based on at least one of reflectivity, pulse width, and echo ID, and the target detection results that do not meet the preset criteria are filtered out.

[0030] In one embodiment of the above target detection method, the method further includes:

[0031] The target detection network, consisting of the three-dimensional feature extraction network, the two-dimensional detection network, and the detection head network, is trained using a point cloud training set.

[0032] The point cloud training set is divided into positive and negative samples based on the IOU-based label allocation strategy.

[0033] In a second aspect, the present invention provides a storage medium adapted to store a plurality of program codes, the program codes being adapted to be loaded and run by a processor to perform the target detection method described in any of the above embodiments.

[0034] In a third aspect, the present invention provides an electronic device including a processor that executes instructions to implement the target detection method described in any of the above embodiments.

[0035] In a fourth aspect, the present invention provides a vehicle comprising the aforementioned electronic equipment.

[0036] By employing the above technical solutions, this invention reduces the dependence of the target detection network on absolute distance and improves the network's anti-interference and generalization capabilities by simultaneously allocating point cloud data in voxel format and normalizing the point cloud data within each voxel. By optimizing the downsampling structure of the sparse 3D convolutional network, the detection range in the altitude direction is expanded without increasing computational cost. Furthermore, by using two binary classification networks to further refine the directional attributes of the heading angle, the heading angle generation method is improved, thus enhancing the accuracy of heading angle detection. Simultaneously, by converting point cloud data based on the point cloud sensor coordinate system to point cloud data based on the application object coordinate system, the method of this invention can be applied to different types of radar sensors, improving its practicality and expanding the data sources available for training the target detection model of this invention. This also helps to further improve the parameters of each network in the model and enhance the accuracy of target detection. Attached Figure Description

[0037] The embodiments of the present invention are described below with reference to the accompanying drawings, in which:

[0038] Figure 1 This is a flowchart of the main steps of the target detection method according to an embodiment of the present invention.

[0039] Figure 2 This is a schematic diagram of the detection head network structure according to an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram showing the positions of the first orientation attribute and the second orientation attribute in an embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0043] First read Figure 1 , Figure 1 This is a flowchart illustrating the main steps of the target detection method according to an embodiment of the present invention. Figure 1 As shown, the target detection method of the present invention includes:

[0044] Step S101: Acquire point cloud data;

[0045] Step S102: Voxelize the point cloud data to obtain voxelized point cloud data;

[0046] Step S103: Based on the voxelized point cloud data, obtain a 3D point cloud feature map through a 3D feature extraction network;

[0047] Step S104: Based on the 3D point cloud feature map, obtain the 2D point cloud feature map through a 2D detection network.

[0048] Step S105: Based on the two-dimensional point cloud feature map, obtain the target category and the three-dimensional detection box containing the heading angle through the detection head network.

[0049] In step S101, preferably, raw point cloud data is acquired using an onboard LiDAR. In an embodiment of the present invention, the raw point cloud data is first transformed into point cloud data based on the LiDAR coordinate system, which is then transformed into point cloud data based on the vehicle coordinate system required in step S102. As an example, when the three-dimensional coordinate directions of the LiDAR coordinate system and the vehicle coordinate system are consistent, the coordinate transformation between the two coordinate systems can be completed simply by translation; when the three-dimensional coordinate directions of the onboard LiDAR coordinate system and the vehicle coordinate system are inconsistent, the coordinate transformation between the two coordinate systems can be completed by rotation and translation.

[0050] Coordinate transformation can broaden the applicability of the method of the present invention. By simply performing the corresponding coordinate transformation according to the positional relationship between the vehicle-mounted lidar and the vehicle, the method of the present invention can be applied to different lidars.

[0051] After completing the coordinate transformation of the point cloud data, in order to further enhance the point cloud features and improve the target detection effect, data augmentation can be performed on the point cloud data according to the actual situation. Point cloud data augmentation methods include at least one of translation, yaw angle rotation, and scaling.

[0052] In step S102, the purpose of voxelizing point cloud data is to convert sparse 3D point clouds into dense 3D voxel features in order to reduce the computational load of data processing.

[0053] First, three-dimensional Cartesian coordinate systems are established for the point cloud data and the voxel space, respectively. In the embodiments of this invention, both the point cloud data and the three-dimensional Cartesian coordinate systems in the voxel space use the vehicle coordinate system. The vehicle coordinate system can be set with the direction of the vehicle's front as the positive x-axis, the left side of the vehicle's front as the positive y-axis, and the top of the vehicle as the positive z-axis. That is, the x-axis and y-axis are two mutually perpendicular coordinate axes in the horizontal plane, and z is a coordinate axis perpendicular to the horizontal plane in the height direction.

[0054] Based on the 3D coordinates and voxel positions of the point cloud points, each point in the point cloud data is assigned to a corresponding voxel. At this time, the number of point cloud points in each voxel is not fixed; it may be one or more, or it may be zero.

[0055] Since the subsequent data processing is performed on a voxel-by-voxel basis, it is necessary to perform preliminary feature extraction on the point cloud points within each voxel to obtain voxelized, regular point cloud data. Specifically, based on the spatial location of each voxel, all point cloud points within each voxel are normalized, including obtaining the voxelized point cloud dimensional features and voxelized point cloud coordinates.

[0056] The method for obtaining the dimensional features of a voxelized point cloud includes: obtaining the mean value of each dimension of all point cloud points within a voxel; subtracting the center position value of the voxel corresponding to that dimension from the mean value of each dimension to obtain the dimensional features of the voxelized point cloud. The method for obtaining the dimensional features of a voxelized point cloud is as follows:

[0057]

[0058]

[0059]

[0060] Among them, X f Y f and Z f These represent the voxelized point cloud dimensionality features in the x, y, and z directions of the f-th voxel, respectively; x i y i and z i X represents the x, y, and z coordinates of the i-th point cloud point within the f-th voxel; voxel_center Y voxel_center and Z voxel_center are the coordinates of the center positions of the f-th voxel in the x, y, and z directions, respectively. , respectively, are the mean values ​​of the i-th point cloud point in the x, y, and z directions within the f-th void; n is the number of point cloud points within the f-th void, and is at least 1.

[0061] The method for obtaining voxelized point cloud coordinates includes: obtaining the mean value of all point cloud points within a voxel for each dimension; obtaining the difference between the mean value of each dimension and the minimum detection range of the corresponding LiDAR, and dividing this difference by the voxel size corresponding to that dimension to obtain the normalized voxelized point cloud coordinates. The method for obtaining voxelized point cloud coordinates is as follows:

[0062]

[0063]

[0064]

[0065] Among them, X coordinate_f Y coordinate_f and Z coordinate_f These are the voxelized point cloud coordinates of the f-th voxel in the x, y, and z directions, respectively. , where are the mean values ​​of the i-th point cloud point in the f-th voxel in the x, y, and z directions, respectively; n is the number of point cloud points in the f-th voxel, and is at least 1; Range_min x Range_min y and Range_min z These are the minimum detection ranges of the lidar in the x, y, and z directions, respectively. and These represent the voxel dimensions in the x, y, and z directions, respectively.

[0066] The detection range in the x, y, and z directions depends on the performance of the lidar. For example, if the lidar's detection range in the x direction is 0–300 meters, then Range_min x =0.

[0067] It should be noted that for voxels with zero points in the point cloud, it is not necessary to calculate the voxelized point cloud dimensional features and voxelized point cloud coordinates, and the voxel does not need to be input into a sparse 3D convolutional network.

[0068] The voxel size is typically set based on actual conditions. Larger voxel sizes result in less computational data but more lost features in the voxelized point cloud data; smaller voxel sizes result in more computational data but more retained features in the voxelized point cloud data. Preferably, in an embodiment of the present invention, the voxel size is:

[0069] and

[0070] It should be noted that when performing voxelization of point cloud dimensional features, the features of all point cloud points within a voxel are normalized by calculating the mean of each dimension. Then, the center position of the corresponding voxel is subtracted from this mean to obtain the distance relative to the center of each voxel, removing the distance relative to the center of the vehicle (equivalent to an onboard LiDAR). The advantage of this approach is that it reduces the dependence of the neural network used to process voxelized point cloud data on absolute distances, improves the neural network's robustness against interference from LiDAR point cloud distributions, and enhances the generalization ability of the neural network.

[0071] For other feature data in the point cloud data, such as heading angle, reflectivity, pulse width, etc., calculate the mean of other feature data of all point cloud points in each voxel to obtain other voxelized feature data, which together serve as the input data of the sparse 3D convolutional network in step S103.

[0072] In step S103, a sparse 3D convolutional network is used to extract features from the voxelized point cloud data to obtain a 3D point cloud feature map. In this embodiment of the invention, the network parameters for downsampling spatial features are optimized, using convolutional layers with a kernel size of 3, a stride of 3, and zero padding to downsample the z-axis (height direction). By optimizing the network parameters, the detection range in the height direction can be increased without changing the computational cost of the network.

[0073] To increase the receptive field and combine more layers of point cloud features, in step S104, the three-dimensional point cloud feature map is input into a two-dimensional detection network for two-dimensional point cloud feature extraction. In an embodiment of the present invention, the three-dimensional point cloud feature map is first compressed in the height direction, and then the compressed three-dimensional point cloud feature map is extracted using a VGG-like network to obtain a two-dimensional point cloud feature map.

[0074] In step S105, the two-dimensional point cloud feature map is input into the detection head network for target classification and localization. The structure of the detection head network is as follows: Figure 2 As shown, the detection head network includes: a convolutional unit 201, a target classification unit 202, a 3D detection box extraction unit 203, a heading angle numerical feature extraction unit 204, a heading angle orientation feature extraction unit 205, a first binary classification unit 206, a second binary classification unit 207, a heading angle generation unit 208, a result filtering unit 209, and a bounding box unit 210.

[0075] The two-dimensional point cloud feature map obtained in step S104 is input into convolutional unit 201 for further feature extraction to obtain an enhanced two-dimensional point cloud feature map. The enhanced two-dimensional point cloud feature map is then input into target classification unit 202, three-dimensional detection box extraction unit 203, heading angle numerical feature extraction unit 204, and heading angle orientation feature extraction unit 205, respectively.

[0076] The output of target classification unit 202 is the target type.

[0077] The output of the 3D detection box extraction unit 203 is a 3D detection box, which includes spatial positions x, y and z, length, width and height values ​​l, w and h, as well as other point cloud attribute features, such as the proportion of high reflectivity point cloud points in the 3D detection box feat1, etc.

[0078] The output of the heading angle numerical feature extraction unit 204 is the heading angle numerical value.

[0079] The output of the heading angle orientation feature extraction unit 205 is the heading angle orientation feature.

[0080] The heading angle orientation feature is input into the first binary classification unit 206 and the second binary classification unit 207 respectively. The output of the first binary classification unit 206 is the first orientation attribute, and the output of the second binary classification unit 207 is the second orientation attribute.

[0081] The heading angle value, the first heading attribute, and the second heading attribute are input into the heading angle generation unit 208 to obtain the heading angle.

[0082] The target classification results, 3D detection box parameters, and heading angle input results are filtered by the filtering unit 209, which uses non-maximum suppression to filter out redundant 3D detection boxes.

[0083] The filtered 3D detection bounding box, target type, and heading angle are input into the bounding box unit to obtain the target detection result, which is represented in bounding box form. As an example, the target detection result can be represented as bbox = (C, x, y, z, l, w, h, yaw, feat1), where C is the target type, such as small motor vehicle, large motor vehicle, pedestrian, etc.; x, y, z are the spatial positions of the 3D detection bounding box; l, w, h are the length, width, and height of the 3D detection bounding box; yaw is the heading angle; and feat1 is the proportion of high reflectivity point cloud points in the 3D detection bounding box.

[0084] In practical applications, radar data often contains interference signals, which can lead to false signals in target detection results. To further improve the accuracy of target detection, other attributes of the point cloud data within the 3D detection box can be analyzed, such as reflectivity, pulse width, and echo ID, to filter out false targets that do not meet preset standards. For example, when there are traffic signs with reflective materials on the road, the reflection intensity of these signs is relatively high, causing the point cloud outline formed by the traffic signs to spread outwards, creating a point cloud shape larger than the actual object. This can easily lead to misidentification as a vehicle in target detection results, affecting vehicle assisted driving control. In this case, the proportion of high-reflectivity point cloud points in the target detection results can be considered. If this proportion exceeds a preset high-reflectivity threshold, such as 60%, the target can be determined to be not a vehicle.

[0085] As an example, in an embodiment of the present invention, the convolutional unit 201 is a two-layer convolutional network, the target classification unit 202, the three-dimensional detection box extraction unit 203, the heading angle numerical feature extraction unit 204, and the heading angle orientation feature extraction unit 205 are single-layer convolutional networks; the first binary classification unit 206 and the second binary classification unit 207 are binary classification networks.

[0086] Next, we will further combine Figure 3 This describes the method used in the heading angle generation unit 208 to generate a heading angle based on the heading angle value, the first heading attribute, and the second heading attribute.

[0087] In the aforementioned vehicle coordinate system, the x-axis and y-axis are two mutually perpendicular coordinate axes in the vehicle's horizontal plane. The direction of the vehicle's front is defined as the positive direction of the x-axis, and the left side of the vehicle's front is defined as the positive direction of the y-axis. The positive x-axis is 0°, with counter-clockwise rotation being positive and clockwise rotation being negative. That is, the angle between the positive x-axis and positive y-axis is +90°, and the angle between the positive x-axis and negative y-axis is -90°. The heading angle ranges from -90° to +90°.

[0088] It should be noted that, in the embodiments of the present invention, the method for representing the range of angle values ​​is as follows: when the range of angle values ​​is represented as [A,B], it means that the range of angle values ​​is between A and B, and includes A and B; when the range of angle values ​​is represented as (A,B), it means that the range of angle values ​​is between A and B, but does not include A and B; when the range of angle values ​​is represented as (A,B], it means that the range of angle values ​​is between A and B, includes B, but does not include A.

[0089] The first orientation attribute includes forward and backward directions, where forward direction corresponds to [-45°, 45°] in the vehicle's horizontal Cartesian coordinate system, and backward direction corresponds to [135°, 180°] and (-180°, -135°) in the vehicle's horizontal Cartesian coordinate system.

[0090] The second orientation attribute includes left and right directions. Left direction corresponds to (+45°, +135°) in the vehicle's horizontal Cartesian coordinate system, and right direction corresponds to (-135°, -45°) in the vehicle's horizontal Cartesian coordinate system.

[0091] By combining the angle range of the heading angle orientation attribute, heading angle values ​​with the same heading angle discrimination method can be divided into different angle ranges, thereby realizing the determination of the heading angle.

[0092] When the heading angle value is within the first angle range [-45°, 45°], only the first heading attribute is valid, and the heading angle is determined based on the first heading attribute. When the first heading attribute is forward, the heading angle is the position corresponding to the heading angle value. When the first heading attribute is backward, the heading angle is the position where the heading angle value is rotated 180° counterclockwise. As an example, if the heading angle value is 30°, and the first heading attribute is forward, the heading angle is +30°; if the first heading attribute is backward, rotating 180° counterclockwise, i.e., 30° plus 180°, the heading angle value is 210°, and the corresponding heading angle is -150°.

[0093] When the heading angle value is within the second angle range (45°, 90°), only the second heading attribute is valid, and the heading angle is determined based on the second heading attribute. When the second heading attribute is left, the heading angle is the position corresponding to the heading angle value. When the second heading attribute is right, the heading angle is the position 180° counterclockwise from the heading angle value. For example, if the heading angle value is 60°, and the second heading attribute is left, the heading angle is +60°; if the second heading attribute is right, rotating 180° counterclockwise (60° plus 180°) gives a heading angle value of 240°, and the corresponding heading angle is -120°.

[0094] When the heading angle value is within the third angle range (-90°, -45°), only the second heading attribute is valid, and the heading angle is determined based on the second heading attribute. When the second heading attribute is left, the heading angle is the position 180° counterclockwise from the heading angle value. When the second heading attribute is right, the heading angle is the position corresponding to the heading angle value. For example, if the heading angle value is -60°, and the second heading attribute is left, rotating it 180° counterclockwise is equivalent to adding 180° to -60°, resulting in a heading angle of 120°, and the corresponding heading angle is +120°. If the second heading attribute is right, the heading angle is -60°.

[0095] When the heading angle value is -90° (first boundary angle) or 90° (second boundary angle), the heading angle is the position corresponding to the heading angle value, that is, the heading angles are -90° or +90° respectively.

[0096] It should be noted that, in the embodiments of the present invention, the heading angle range is (-180°, +180°).

[0097] In this embodiment of the invention, during point cloud data processing, the directional attribute of the heading angle is further divided into forward, backward, left, and right directions, and the boundary line for dividing the heading angle direction attribute is the angle bisector of the x-axis and y-axis directions. Combining the boundary line of the heading angle direction attribute with the x-axis and y-axis directions, the direction of the heading angle is divided into 8 intervals in the horizontal plane, thereby making the detection of the heading angle more accurate.

[0098] Before using the target detection network, which is mainly composed of a sparse 3D convolutional network, a two-dimensional detection network, and a detection head network, the target detection network needs to be trained as a whole using a point cloud training set.

[0099] Similarly, the coordinates of the point cloud data in the point cloud training set are also coordinate-transformed and relative to the vehicle. Through coordinate transformation, data from different LiDAR systems can be used, expanding the sources of the point cloud training dataset.

[0100] When generating point cloud labels, positive and negative samples in the point cloud training set are divided based on the IOU label allocation strategy, and the object detection network is trained by forward and backward propagation using the TensorFlow deep learning framework.

[0101] It should be noted that sparse 3D convolutional networks, VGG-like networks, nonmaximum suppression, and the TensorFlow deep learning framework are all commonly used feature extraction and data filtering methods by those skilled in the art, and will not be elaborated upon here.

[0102] Furthermore, the present invention also provides a storage medium that can be configured to store a program for executing the target detection method of the above-described method embodiments. This program can be loaded and run by a processor to implement the target detection method. For ease of explanation, only the parts related to the embodiments of the present invention are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of the present invention. The storage medium can be a storage device comprising various electronic devices. Optionally, in the embodiments of the present invention, the storage medium is a non-transitory read / write storage medium.

[0103] Furthermore, the present invention also provides an electronic device including a processor, which can be configured to execute instructions to implement the target detection method of the above-described method embodiments. For ease of explanation, only the parts related to the embodiments of the present invention are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of the present invention. The target detection device may be a control device comprising various electronic devices.

[0104] Furthermore, the present invention also provides a vehicle including the aforementioned electronic device, which includes a processor configured to execute instructions to implement the target detection method of the above-described method embodiments. Optionally, the vehicle is a new energy vehicle equipped with an onboard LiDAR and possessing driver assistance functions.

[0105] Those skilled in the art will recognize that the method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the invention.

[0106] It should be noted that the terms "first," "second," and other ordinal numbers in the specification, claims, and accompanying drawings of this invention are used only to distinguish similar objects and not to describe or indicate a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein.

[0107] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A target detection method, characterized in that, The method includes: Acquire point cloud data; The point cloud data is voxelized to obtain voxelized point cloud data; Based on the voxelized point cloud data, a three-dimensional point cloud feature map is obtained through a three-dimensional feature extraction network, wherein the three-dimensional feature extraction network is a sparse 3D convolutional network. Based on the three-dimensional point cloud feature map, a two-dimensional point cloud feature map is obtained through a two-dimensional detection network; Based on the two-dimensional point cloud feature map, the target category and a three-dimensional detection box including the heading angle are obtained through the detection head network; The heading angle is generated based on a heading angle value and a heading angle orientation attribute. The heading angle orientation attribute includes a first orientation attribute and a second orientation attribute. The first orientation attribute includes forward and backward directions, and the second orientation attribute includes left and right directions. Specifically, it includes: When the heading angle value is within the first angle range, the heading angle is obtained according to the first orientation attribute; When the heading angle value is within the second angle range, the heading angle is obtained according to the second orientation attribute; When the heading angle value is within the third angle range, the heading angle is obtained according to the second orientation attribute; When the heading angle value is the first boundary angle or the second boundary angle, the heading angle is the position corresponding to the heading angle value; The heading angle is a value in a Cartesian coordinate system (x, y), with the heading angle in the positive x-axis direction being 0, clockwise rotation being negative, and counterclockwise rotation being positive. The first angle range is [-45°, 45°], the second angle range is (45°, 90°), and the third angle range is (-90°, -45°). The first boundary angle is -90°, and the second boundary angle is 90°.

2. The target detection method according to claim 1, characterized in that, Based on the two-dimensional point cloud feature map, the target category and the three-dimensional detection box including the heading angle are obtained through the detection head network, including: The target category, the three-dimensional detection box, the heading angle value, and the heading angle orientation feature are obtained from the two-dimensional point cloud feature map. The heading angle orientation attribute is obtained based on the heading angle orientation feature.

3. The target detection method according to claim 1, characterized in that, The point cloud data is voxelized to obtain voxelized point cloud data, including: Obtain the mean coordinates of each dimension of all point cloud points within each voxel; Subtracting the coordinates of the corresponding dimension of the center position of the voxel from the mean of the coordinates yields the voxelized point cloud dimensional features. The difference between the mean coordinate of each dimension and the minimum radar detection range corresponding to that dimension is obtained. The difference is then divided by the voxel size of the corresponding dimension to obtain the voxelized point cloud coordinates.

4. The target detection method according to claim 1, characterized in that, The sparse 3D convolutional network downsamples in the height dimension using convolutional layers with a kernel size of 3, a stride of 3, and padding of 0.

5. The target detection method according to claim 1, characterized in that, The method further includes: The target detection results are analyzed based on at least one of reflectivity, pulse width, and echo ID, and the target detection results that do not meet the preset criteria are filtered out.

6. The target detection method according to any one of claims 1-5, characterized in that, The method further includes: The target detection network, consisting of the three-dimensional feature extraction network, the two-dimensional detection network, and the detection head network, is trained using a point cloud training set. The point cloud training set is divided into positive and negative samples based on the IOU-based label allocation strategy.

7. A storage medium suitable for storing multiple lines of program code, characterized in that, The program code is adapted to be loaded and run by a processor to perform the target detection method according to any one of claims 1 to 6.

8. An electronic device, comprising a processor, characterized in that, The processor executes instructions to implement the target detection method according to any one of claims 1 to 6.

9. A vehicle, characterized in that, The vehicle includes the electronic equipment as described in claim 8.

Images