True value generation method, detection method and network training method for drivable area

Abstract

The application discloses a true value generation method, a detection method and a network training method of a drivable area, and belongs to the technical field of automatic driving. The true value generation method comprises the following steps: obtaining labeled information of a static map comprising obstacle contours, the static map being constructed based on multi-path surround video data, laser radar data and positioning data; obtaining a plurality of image groups, each image group comprising a plurality of surround images simultaneously collected in the multi-path surround video data; determining coordinate points of each obstacle contour in a self-vehicle coordinate system based on pose information in the laser radar data matched with a target timestamp of a target image group; determining the drivable area under different visual angles by taking the installation position of the multi-path surround camera as an observation point by using a ray casting method; taking the intersection of each drivable area as a target drivable area, and generating a true value result corresponding to the target image group based on target boundary points of the target drivable area in a bird's-eye view angle coordinate system. The application can significantly improve the true value labeling efficiency.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, specifically to methods for generating truth values, detecting, and training networks for drivable areas. Background Technology

[0002] In autonomous parking scenarios, drivable area identification is a key technology for achieving accurate path planning and autonomous parking. Related methods rely on high-precision sensors and complex trajectory correction algorithms, which, while enabling environmental perception, still suffer from limitations such as low annotation efficiency. Summary of the Invention

[0003] This application provides a method for generating ground truth values, a method for detecting drivable areas, and a method for training a network, aiming to solve the problem of low annotation efficiency in related drivable area identification methods.

[0004] Firstly, a method for generating truth values ​​for drivable regions is provided, including: Obtain annotation information for the static map. The static map is constructed based on multi-channel surround view video data, LiDAR data, and positioning data collected by the ground truth vehicle. The annotation information includes obstacle outlines. Multiple image groups are acquired by synchronously extracting frames from multi-channel surround view video data. Each image group includes multi-channel surround view images acquired simultaneously. For each of the above multiple image groups, based on the pose information in the lidar data that matches the target timestamp of the target image group, the coordinate points of each obstacle contour in the vehicle coordinate system are determined. In the vehicle coordinate system, the ray projection method is used, with the installation position of the multi-channel surround view camera corresponding to the above multi-channel surround view video data as the observation point. Based on the above coordinate points within the preset visual range, the drivable area under different angles of the multi-channel surround view camera is determined. The intersection of each drivable area is taken as the target drivable area corresponding to the target image group. The target boundary points of the target drivable area in the bird's-eye view coordinate system are determined, and the true value results indicating each target boundary point are generated for the target image group.

[0005] In some embodiments, after taking the intersection of each drivable region as the target drivable region corresponding to the target image group, the method further includes: determining a category label corresponding to each boundary point of the target drivable region based on the obstacle contour, wherein the category label is used to characterize the obstacle boundary or extension line. The true values ​​also indicate the category label corresponding to each target boundary point.

[0006] In some embodiments, acquiring multiple image groups includes: Extract multiple frames of surround view images from one surround view video stream; For each first image in a multi-frame surround view image, based on the first timestamp of the first image, a second image is selected from each surround view video except for one surround view video in the multi-channel surround view video data, and the first image and each second image are combined into an image group, with the target timestamp of the image group being the first timestamp; wherein, the difference between the second timestamp of the second image and the first timestamp is less than or equal to the time difference threshold.

[0007] In some embodiments, the pose information that matches the target timestamp of the target image group includes the first pose information corresponding to the two timestamps with the smallest difference from the target timestamp in the lidar data. Based on the pose information matching the timestamps of the target image group in the LiDAR data, the coordinates of each obstacle contour in the vehicle coordinate system are determined, including: Interpolate each first pose information to obtain the second pose information corresponding to the target timestamp; Based on the second pose information, the coordinates of each obstacle contour in the vehicle coordinate system are determined.

[0008] In some embodiments, the above generation method further includes: Filter out coordinate points that are outside the preset visible range from each coordinate point; Using a light projection method, with the installation position of the multi-channel surround-view camera corresponding to the multi-channel surround-view video data as the observation point, and based on the coordinates of the points within the preset field of view, the drivable area under different viewing angles of the multi-channel surround-view camera is determined, including: Extract vertices from the boundaries of the preset visible area and from each coordinate point within the preset visible area; For each target camera in a multi-view camera system, sort the vertices according to their polar angles relative to the observation point of the target camera. Rays are emitted from the observation point to each vertex in the order listed above. Detect the nearest occlusion point of each ray, and determine the drivable area from the perspective of the target camera based on the nearest occlusion point of each ray.

[0009] Secondly, a method for detecting drivable areas is provided, including: The first feature map of each of the acquired multi-path surround view images is extracted using the feature extraction network in the pre-trained drivable area detection network; the drivable area detection network also includes a view transformation network and a segmentation head network. Using a view transformation network, each first feature map is transformed into a second feature map in the bird's-eye view space; The target feature map of the multi-path loop view image is determined based on each second feature map; Using a segmentation head network, drivable region detection is performed based on the target feature map, and detection results are generated. The detection results indicate the boundary points of the drivable region corresponding to the multi-path surround view image.

[0010] In some embodiments, a feature extraction network in a pre-trained drivable area detection network is used to extract a first feature map of each of the acquired multi-path surround view images, including: A feature extraction network is used to extract multi-scale features from each loop image, and the multi-scale features are fused into a first feature map.

[0011] In some embodiments, the feature extraction network includes a fast feature network and a feature pyramid network corresponding to the multi-path loop images respectively; Using a feature extraction network, multi-scale features of each loop image are extracted and fused into a first feature map, including: The fast feature network corresponding to each loop image is used to extract multi-scale features of each loop image, and the feature pyramid network corresponding to each loop image is used to fuse the multi-scale features into a first feature map.

[0012] In some embodiments, the segmentation head network is a three-channel segmentation head network, and the detection results also include category labels for boundary points, which are used to characterize obstacle boundaries or extension lines.

[0013] Thirdly, a training method for a drivable area detection network is provided, including: Obtain the training dataset, in which each training data includes an image group and the corresponding ground truth result. The image group includes multiple surround view images acquired simultaneously, and the ground truth result indicates the target boundary point of the target drivable area corresponding to the image group in the bird's-eye view coordinate system. The feature extraction network in the drivable region detection network is used to extract the first feature map of each around view image in the target image group in the training dataset; the drivable region detection network also includes a viewpoint transformation network and a segmentation head network; Using a view transformation network, each first feature map is transformed into a second feature map in the bird's-eye view space; The target feature map of the target image group is determined based on each second feature map; Using a segmentation head network, drivable regions are detected based on target feature maps, and detection results are generated. The detection results indicate the boundary points of the drivable regions corresponding to the target image group. The parameters of the drivable area detection network are adjusted with the goal of minimizing the difference between the ground truth and the detection results corresponding to the target image group.

[0014] Beneficial effects: The solution provided in this application involves acquiring annotation information for a static map. This static map is constructed based on multi-channel surround-view video data, LiDAR data, and positioning data collected by a ground truth vehicle. The annotation information includes obstacle outlines. Next, multiple image groups are acquired. These image groups are obtained by synchronously extracting frames from the multi-channel surround-view video data. Each image group includes simultaneously acquired multi-channel surround-view images. Then, for each target image group within these multiple image groups, the coordinates of each obstacle outline in the vehicle's coordinate system are determined based on the pose information in the LiDAR data that matches the target timestamp of the target image group. Next, in the vehicle's coordinate system, using a ray casting method, with the installation position of the multi-channel surround-view camera corresponding to the aforementioned multi-channel surround-view video data as the observation point, and based on the coordinates within a preset visible range, the drivable area under different viewing angles of the multi-channel surround-view camera is determined. The intersection of these drivable areas is then used as the target drivable area corresponding to the target image group. The target boundary points of the target drivable area in the bird's-eye view coordinate system are determined, and ground truth results indicating each target boundary point are generated for the target image group. This solution proposes a semi-automatic annotation method based on static high-precision maps. By using spatiotemporal alignment, map annotations are mapped to BEV images, significantly improving annotation efficiency compared to single-frame annotation methods, while also achieving higher ground truth accuracy. Using multiple image sets and corresponding ground truth results to train a drivable area detection network can improve the accuracy of drivable area recognition in complex scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the truth generation method for drivable areas provided in an embodiment of this application; Figure 2 This is a schematic diagram of obtaining a drivable area using a light projection method provided in an embodiment of this application; Figure 3 This is a flowchart illustrating the training method for the drivable area detection network provided in an embodiment of this application; Figure 4 This is a schematic diagram of the drivable area detection network provided in the embodiments of this application; Figure 5 This is a flowchart illustrating the method for detecting drivable areas provided in the embodiments of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0019] "A and / or B" includes the following three combinations: A only, B only, and a combination of A and B.

[0020] The use of "applies to" or "configured to" in this application implies open and inclusive language, which does not exclude the applicability to or configuration to devices performing additional tasks or steps. Additionally, the use of "based on" implies openness and inclusivity, because processes, steps, calculations, or other actions "based on" one or more conditions or values ​​may in practice be based on additional conditions or values ​​beyond those stated.

[0021] In this application, the term "exemplary" is used to mean "used as an example, illustration, or description." Any embodiment described as "exemplary" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes are not described in detail to avoid obscuring the description of this application with unnecessary detail. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0022] In autonomous parking scenarios, drivable area identification is a key technology for achieving accurate path planning and autonomous parking. Related technologies rely on high-precision sensors and complex trajectory correction algorithms. While they can achieve environmental perception, limitations remain, including low annotation efficiency, insufficient dynamic adaptability, and high computational complexity.

[0023] One related technique proposes a parking drivable area detection method based on cascaded forests and stacked ensemble learning. This method extracts features and outputs multiple segmentation results through multi-layered cascaded tree-structured base learners, which are then fused through stacked ensemble learning to obtain the final detection result. However, the technical characteristics of this technique inherently limit its capabilities: traditional machine learning architectures have limited feature representation capabilities, making it difficult to handle complex and varied parking scenarios; reliance on manual feature engineering leads to insufficient generalization ability; and while the multi-layered cascaded structure improves model stability, it sacrifices computational efficiency. These technical characteristics collectively result in objective defects such as limited detection accuracy, significant computational latency, and insufficient scene adaptability.

[0024] Another related technology proposes a drivable area detection method based on laser point clouds. Although the method of using laser point clouds to detect drivable areas can cope with inclement weather to a certain extent, it still has inherent drawbacks such as limited environmental adaptability, insufficient spatial resolution, weak dynamic obstacle recognition capability, and high system cost.

[0025] In view of this, embodiments of this application provide a ground truth generation method, a detection method, and a network training method for drivable areas. Among them, a semi-automatic annotation method based on static high-precision maps is proposed. By aligning the map annotations to the BEV image through spatiotemporal alignment, the annotation efficiency is significantly improved compared to the single-frame annotation method, and the accuracy of the ground truth is higher. Using multiple image groups and corresponding ground truth results from this ground truth generation method to train the drivable area detection network can significantly improve the accuracy of drivable area recognition in complex scenarios, thereby solving at least one of the above-mentioned technical problems.

[0026] In some embodiments, the truth generation method, detection method, and network training method for drivable areas in this application can be executed by any platform, system, server, device, device cluster, or apparatus with computing and processing capabilities.

[0027] In some embodiments, the method for detecting drivable areas in this application may be performed by an autonomous driving system. Furthermore, the autonomous driving system may include, but is not limited to, an autonomous parking system.

[0028] Figure 1 This is a flowchart illustrating the truth value generation method for drivable areas provided in an embodiment of this application. For example... Figure 1 As shown, this truth value generation method includes the following steps: S101: Obtain the annotation information of the static map. The static map is constructed based on the multi-channel surround view video data, LiDAR data and positioning data collected by the ground truth vehicle. The annotation information includes the outline of obstacles. S103: Acquire multiple image groups, which are obtained by synchronously extracting frames from multi-channel surround view video data. Each image group includes multi-channel surround view images acquired simultaneously. S105: For each target image group in the above multiple image groups, based on the pose information in the lidar data that matches the target timestamp of the target image group, determine the coordinate points of each obstacle contour in the vehicle coordinate system. S107: In the vehicle coordinate system, using the ray projection method, with the installation position of the multi-channel surround view camera corresponding to the above multi-channel surround view video data as the observation point, based on the above coordinate points within the preset visible range, the drivable area under different angles of the multi-channel surround view camera is determined. S109: Take the intersection of each drivable area as the target drivable area corresponding to the target image group, determine the target boundary points of the target drivable area in the bird's-eye view coordinate system, and generate the true value results corresponding to the target image group that indicate each target boundary point.

[0029] Figure 1 The corresponding implementation provides a semi-automatic annotation method based on static high-precision maps. By aligning the map annotations to the BEV image through spatiotemporal alignment, the annotation efficiency is significantly improved compared to the single-frame annotation method, while the accuracy of the ground truth is also higher.

[0030] Steps S101 to S109 will be explained below.

[0031] In step S101, the annotation information of the static map is obtained. The static map is constructed based on multi-channel surround-view video data, LiDAR data, and positioning data collected by the ground truth vehicle. The annotation information includes obstacle outlines. The ground truth vehicle is equipped with multiple surround-view cameras, which collect multi-channel surround-view video data. In one example, the "multi-channel" in this application refers to four channels, and the multi-channel surround-view cameras may include front, left, rear, and right surround-view fisheye cameras. The LiDAR data may include point cloud information and pose information. The annotation information of the static map can be obtained by annotators who annotate obstacle outlines on the static map based on the point cloud information. The coordinates of the obstacle outlines in this annotation information are coordinates in the global coordinate system.

[0032] In step S103, multiple image groups are acquired by synchronously extracting frames from multi-channel surround view video data. Each image group includes multi-channel surround view images acquired simultaneously.

[0033] In practice, both multi-channel surround-view video data and LiDAR data include timestamp information. Acquiring multiple image groups involves: extracting multiple frames of surround-view images from one surround-view video stream; for each frame of the first image, based on the first timestamp of the first image, selecting second images from all other surround-view videos in the multi-channel surround-view video data, and combining the first image and each second image into an image group, with the target timestamp of the image group being the first timestamp; wherein the difference between the second timestamp of the second image and the first timestamp is less than or equal to a time difference threshold. Therefore, time alignment between multiple surround-view images can be achieved using the timestamp of the single surround-view video stream as a reference.

[0034] It should be noted that the aforementioned surround-view video can be any one of the multiple surround-view video data, or a forward-looking surround-view video. The time difference threshold can be any value among 1ms, 2ms, 3ms, 4ms, and 5ms, or any value within a range formed by any two values; no specific limitation is made here.

[0035] In step S105, each of the above multiple image groups is taken as a target image group, and the coordinates of each obstacle contour in the vehicle coordinate system are determined based on the pose information in the lidar data that matches the target timestamp of the target image group.

[0036] As one implementation method, the pose information matching the target timestamp of the target image group can include the first pose information corresponding to the two timestamps in the LiDAR data with the smallest difference from the target timestamp (one greater than the target timestamp and the other less than the target timestamp). Based on the pose information matching the target timestamp of the target image group in the LiDAR data, the coordinate points of each obstacle contour in the vehicle coordinate system are determined, including: interpolating each first pose information to obtain the second pose information corresponding to the target timestamp; and based on the second pose information, determining the coordinate points of each obstacle contour in the vehicle coordinate system.

[0037] The second pose information is the pose of the LiDAR relative to the origin of the global coordinate system (defined as the device's position when powered on). Based on this pose, the relationship between the LiDAR and the global coordinate system at the target timestamp can be determined, i.e., `lidar2global`. In `lidar2global`, `lidar` represents the LiDAR, `2` indicates relative to, and `global` represents the global coordinate system. From the calibration parameters of the surround-view camera and the LiDAR, the extrinsic parameters of the LiDAR relative to the vehicle can be obtained, i.e., `lidar2ego` (the relationship between the LiDAR and the vehicle's coordinate system). Combining `lidar2global`, the relationship between the vehicle and the global coordinate system can be obtained, i.e., `ego2global`. Here, `ego` represents the vehicle's coordinate system, with the rear axle center as the origin, the vehicle's forward direction as the positive x-axis, and the left side of the vehicle as the positive y-axis. The coordinates of the obstacle contours in the annotation information are points in the global coordinate system; using `ego2global`, the coordinates of the obstacle contours can be transformed to the vehicle's coordinate system.

[0038] After determining the coordinates of each obstacle outline in the vehicle coordinate system at the target timestamp, coordinate points outside the preset visibility range can be filtered out, and then step S107 can be executed. The preset visibility range can be designed according to the actual scenario. For example, in a parking scenario, the range of x-coordinate values ​​in the preset visibility range can be [-9.84m, 13.20m], and the range of y-coordinate values ​​can be [-10.24m, 10.24m].

[0039] In step S107, in the vehicle coordinate system, the ray projection method is used, with the installation position of the multi-channel surround view camera corresponding to the above-mentioned multi-channel surround view video data as the observation point, and the drivable area under different angles of the multi-channel surround view camera is determined based on the above-mentioned coordinate points within the preset visible range.

[0040] in, Figure 2 This is a schematic diagram illustrating the method of obtaining a drivable area using light projection, as provided in an embodiment of this application. Figure 2 As shown, gray triangles and trapezoids represent obstacles, red represents the light source point (observation point), and blue represents the visible area obtained by rays emitted from the light source point, which is the drivable area in this application. In practice, the core principle of obtaining the drivable area polygon is based on ray projection and angle scanning algorithms, combined with geometric intersection to efficiently determine the visible area originating from the light source point.

[0041] Specifically, using the ray projection method, with the installation position of the multi-channel surround-view camera corresponding to the aforementioned multi-channel surround-view video data as the observation point, and based on the aforementioned coordinate points within the preset visual range, the drivable area under different viewing angles of the multi-channel surround-view camera is determined, including: extracting vertices from the boundary of the preset visual range and each coordinate point within the preset visual range; for each target camera in the multi-channel surround-view camera, sorting each vertex according to the polar angle (counterclockwise direction) relative to the observation point corresponding to the target camera; emitting rays from the observation point to each vertex in the sorted order; detecting the nearest occlusion point of each ray, and determining the drivable area under the viewpoint of the target camera based on the nearest occlusion point of each ray.

[0042] It should be noted that the vertices extracted from the boundaries of the preset visible area and the coordinate points within that area are deduplicated to avoid redundant calculations. Sorting the vertices by their polar angle relative to the target camera's observation point facilitates subsequent scanning. Specifically, when sorting vertices by their polar angle relative to the target camera's observation point, if the polar angles are the same, they are sorted by distance (closest to the nearest). The nearest occlusion point for each ray is the farthest visible intersection point of that ray with all polygon edges in the environment.

[0043] In some embodiments, based on the ray-line segment intersection algorithm, the intersection point detection of a ray and a line segment is based on the following geometric calculations: Ray representation: A ray originates from the starting point. and direction vector The parametric equation is defined as follows: ; in, , This represents the coordinates of any point on the ray, determined by the value of t. This can represent the coordinates of the target camera's installation location in the vehicle's coordinate system. This represents the x-coordinate value of that point. This represents the y-coordinate value at that point. Direction vector. This indicates the direction of ray propagation, where dx is the x-axis component and dy is the y-axis component.

[0044] Line segment representation: The edges of objects in the environment are represented by line segment AB, with the parametric equation as follows: ; in, , This represents the coordinates of any point on the line segment, determined by the value of s. , Let A be the coordinate value of the endpoint A of the line segment in the vehicle coordinate system. , Let B be the coordinate value of the endpoint B of the line segment in the vehicle coordinate system.

[0045] Establish a system of equations to find the intersection points of the ray and the line segment: ; Calculate the determinant: ; According to Cramer's rule, when D≠0, the solution is: ; .

[0046] When selecting the nearest intersection point, retain the intersection point that is closest to the light source (observation point).

[0047] Connect all valid intersections (including the intersections of vertices and rays with edges) to form a closed polygon, which is the drivable region polygon.

[0048] In step S109, the intersection of each drivable area is taken as the target drivable area corresponding to the target image group. The target boundary points of the target drivable area in the bird's-eye view (BEV) coordinate system are determined, and the ground truth results indicating each target boundary point corresponding to the target image group are generated. It should be noted that points in the vehicle coordinate system can be transformed to the BEV coordinate system according to the BEV range parameter, thereby extracting the ground truth results under the target timestamp. The BEV range parameter can be the mapping result of the preset visible range in the BEV coordinate system as described above.

[0049] In some embodiments, after taking the intersection of each drivable region as the target drivable region corresponding to the target image group, the method further includes: determining a category label corresponding to each boundary point of the target drivable region based on the obstacle contour. This category label is used to characterize the obstacle boundary or extension line. The ground truth result also indicates the category label corresponding to each target boundary point. The extension line can be understood as a virtual boundary line formed by ray projection from the camera viewpoint to the edge of a preset visible range.

[0050] In some embodiments, the plurality of image groups and their corresponding ground truth results are stored together. For example, each image group and its corresponding ground truth result can be combined into training data, and the training data can be combined into a training dataset for storage.

[0051] The truth value generation method for drivable areas provided in this application has the following beneficial effects: High annotation efficiency: Only the obstacle outline needs to be annotated on a static map to automatically generate the ground truth value from the BEV perspective in a single frame; Reliable annotation accuracy: Spatiotemporal alignment between static maps and sensor data ensures annotation accuracy and reduces errors caused by manual annotation of a single frame; Highly scalable: Provides high-quality ground truth data support for training autonomous driving perception algorithms.

[0052] Compared with traditional frame-by-frame image annotation methods, the embodiments of this application can significantly improve annotation efficiency while ensuring annotation accuracy, providing a reliable data foundation for the research and development of autonomous driving systems.

[0053] Figure 3 This is a flowchart illustrating the training method for the drivable area detection network provided in an embodiment of this application. Figure 3 As shown, the training method includes the following steps: S301: Obtain the training dataset, where each training data includes an image group and the ground truth result corresponding to the image group. The image group includes multi-path surround view images acquired simultaneously. The ground truth result indicates the target boundary point of the target drivable area corresponding to the image group in the bird's-eye view coordinate system. S303: Utilize the feature extraction network in the drivable region detection network to extract the first feature map of each around-view image of the target image group in the training dataset; wherein, the drivable region detection network also includes a viewpoint transformation network and a segmentation head network; S305: Using a view transformation network, each first feature map is converted into a second feature map in the bird's-eye view space; S307: Determine the target feature map of the target image group based on each second feature map; S309: Using a segmentation head network, drivable region detection is performed based on the target feature map, and detection results are generated. The detection results indicate the boundary points of the drivable region corresponding to the target image group. S311: Adjust the parameters of the drivable area detection network with the goal of minimizing the difference between the ground truth and the detection results corresponding to the target image group.

[0054] Figure 3The corresponding implementation provides a solution that, by enabling the ground truth results to indicate the target boundary points of the target drivable area corresponding to the image group in the bird's-eye view coordinate system, and by utilizing the feature extraction network in the drivable area detection network to extract the first feature map of each surround view image in the training dataset, and by using the view transformation network in the drivable area detection network to transform each first feature map into a second feature map in the bird's-eye view space, and by determining the target feature map of the target image group based on each second feature map, can integrate scattered multi-view image features into a unified BEV feature representation, eliminating feature differences and spatial misalignment problems between different viewpoints. This unified representation method allows the model to directly learn the global environmental relationships under the bird's-eye view, adapting to complex spatial layouts in parking scenarios (such as narrow passages and multiple obstacles), and improving the model's adaptability to complex scenarios. Thus, more accurate environmental perception can be achieved. Moreover, using an end-to-end learning approach instead of manual feature engineering can significantly improve the model's generalization ability.

[0055] The following is a comparison of steps S301 to S311.

[0056] In step S301, a training dataset is obtained, wherein each training data point includes an image group and the corresponding ground truth results. The image group includes multiple simultaneously acquired surround view images, and the ground truth results indicate the target boundary points of the drivable area corresponding to the image group in the bird's-eye view coordinate system. For an explanation of the image group and its corresponding ground truth results, please refer to the relevant explanations above; they will not be repeated here.

[0057] In some embodiments, after obtaining the training dataset, to optimize the model's generalization ability to scenes such as shadows, glare, and abrupt changes in ground texture, random data augmentation can be performed on a small subset of images in the training dataset, including brightness, contrast, and saturation enhancement; shadows and glare can be randomly simulated. Then, step S303 can be executed. Comparative experiments have verified that adding data augmentation can reduce false detections and improve model generalization.

[0058] In step S303, the feature extraction network in the drivable region detection network is used to extract the first feature map of each loop view image of the target image group in the training dataset; wherein, the drivable region detection network also includes a viewpoint transformation network and a segmentation head network.

[0059] The drivable area detection network is a deep neural network used to detect drivable areas. This deep neural network includes a feature extraction network, a viewpoint transformation network, and a segmentation head network.

[0060] When using a feature extraction network to extract the first feature map of each panorama of a target image group, for example, a feature extraction network can be used to extract multi-scale features of each panorama and fuse the multi-scale features into the first feature map.

[0061] Furthermore, the feature extraction network includes a fast feature network (FFNet) and a feature pyramid network (FPN) corresponding to each of the multiple around-view images. FFNet serves as the backbone network in the drivable region detection network. By using FFNet as the backbone network to extract multi-scale features, both model time and performance can be optimized. Figure 4 As shown, multi-scale features of each loop image can be extracted using the FFNet corresponding to each loop image, and the multi-scale features can be fused into a first feature map using the FPN corresponding to each loop image. Specifically, the FPN can perform convolution, upsampling, and feature fusion operations on the multi-scale features respectively, generating a first feature map that combines detailed and semantic information. Figure 4 This is a schematic diagram of the drivable area detection network provided in the embodiments of this application. Figure 4 The multi-view surround view image is shown, including the front view surround view image, the left view surround view image, the rear view surround view image, and the right view surround view image.

[0062] In step S305, a view transformation network is used to convert each first feature map into a second feature map in the bird's-eye view space. Further, as... Figure 4 As shown, the view transformation network can include view transformation sub-networks corresponding to multiple surround view images. Each surround view image's first feature map can be transformed into a second feature map in the bird's-eye view space using its corresponding view transformation sub-network. For example, the view transformation sub-network can be used to sample (e.g., grid sampling) and padding operations on the first feature map, and then the first feature map can be transformed to the BEV space using a projection transformation matrix. Figure 4 The diagram schematically illustrates the view transformation subnetworks FrontBEV, LeftBEV, BackBEV, and RightBEV corresponding to the front-view panoramic image, left-view panoramic image, back-view panoramic image, and right-view panoramic image.

[0063] In step S307, the target feature map of the target image group is determined based on each second feature map. For example, the second feature maps are stitched together and spatially aligned in the BEV space to generate a complete target feature map.

[0064] Furthermore, the drivable area detection network can also include a feature enhancement network. When determining the target feature map of the target image group based on each second feature map, the feature enhancement network can be used to enhance the fused features of each second feature map to obtain a third feature map, and the target feature map can be determined based on the third feature map. Even further, such as... Figure 4 As shown, the feature enhancement network includes FFNet and FPN, with FFNet serving as a secondary backbone network for the fused features (such as...). Figure 4 The BEVFeat shown in the figure is used for depth extraction to obtain multi-scale BEV features. This FPN is used to fuse these multi-scale BEV features into a third feature map.

[0065] Furthermore, such as Figure 4 As shown, the drivable area detection network can also include a U-Net. When determining the target feature map based on the third feature map, the U-Net can be used to process the third feature map to obtain the target feature map. The target feature map has the same size as the ground truth result corresponding to the target image group. In one example, the target feature map is a high-precision feature map with a spatial resolution of 288×256, where each pixel corresponds to an 8cm×8cm region in the actual scene.

[0066] In step S309, the segmentation head network (such as...) is used. Figure 4 The Head shown in the image is used to detect drivable regions based on the target feature map, and the detection results indicate the boundary points of the drivable regions corresponding to the target image group.

[0067] Furthermore, the ground truth results also include the category labels corresponding to the target boundary points of the drivable area in the bird's-eye view coordinate system. These category labels are used to characterize obstacle boundaries or extension lines. The segmentation head network is a three-channel segmentation head network, and the detection results also indicate the category labels of the boundary points of the drivable area corresponding to the target image group. These category labels are used to characterize obstacle boundaries or extension lines.

[0068] For example, the segmentation head network includes a first segmentation head sub-network, a second segmentation head sub-network, and a third segmentation head network. The first segmentation head sub-network is used to detect drivable regions based on the target feature map, generating first detection information, which includes the probability that each pixel in the target feature map belongs to a drivable region. The second segmentation head network is used to detect obstacle boundaries based on the target feature map, generating second detection information, which includes the probability that each pixel in the target feature map belongs to an obstacle boundary. The third segmentation head network is used to detect extension lines based on the target feature map, generating third detection information, which includes the probability that each pixel in the target feature map belongs to an extension line. Based on the first, second, and third detection information, detection results corresponding to the target image group can be generated. In one example, this detection result can be called a semantic segmentation result.

[0069] In step S311, the parameters of the drivable area detection network are adjusted with the goal of minimizing the difference between the ground truth result and the detection result corresponding to the target image group.

[0070] In some embodiments, for learning drivable areas, due to the class imbalance problem between the background (non-drivable areas) and the foreground (drivable areas), this application adds Dice Loss (overlap loss) on top of Focal Loss.

[0071] Furthermore, the ground truth results corresponding to the target image group include the first annotation result, which indicates the target boundary points of the drivable area in the bird's-eye view coordinate system. The detection results corresponding to the target image group include the first detection information as described above. In step S311, the focal loss (Focal Loss) can be determined based on the first detection information, and the overlap loss (Dice Loss) can be determined based on the first annotation result and the first detection information. For example, the overlap loss (Dice Loss) can be determined based on the binary mask (a matrix containing only 0s and 1s, where 1 indicates that a pixel belongs to the drivable area and 0 indicates that a pixel does not belong to the drivable area) of the first annotation result and the first detection information, and the detection loss of the drivable area is set as Loss = Focal Loss + Dice Loss. Then, with the goal of minimizing the training loss, the parameters of the drivable area detection network are adjusted, and the training loss includes the detection loss (Loss).

[0072] It should be noted that the Focal Loss function is an improved loss function proposed to address the class imbalance problem in drivable area recognition tasks. It introduces a dynamic scaling factor on top of the standard cross-entropy, reducing the weight of easily classified samples and increasing the weight of difficult-to-classify samples. This allows the model to focus more on learning samples that are difficult to distinguish, effectively solving the problem of imbalanced background and foreground sample numbers. The Focal Loss is defined as follows: ; in, This can represent the first detection information. It is a balancing factor used to balance the ratio of positive to negative samples. This indicates the focus function.

[0073] Dice Loss directly optimizes overlapping regions, is insensitive to class ratios, and is used to calculate the similarity between predicted and ground truth values. Dice Loss is defined as follows: ; Where X represents the true value (as in the first annotation result above), Y represents the predicted value (as in the binarization mask above), |X∩Y| represents the intersection of X and Y, and |X| and |Y| represent the number of elements in X and Y respectively.

[0074] In some embodiments, the ground truth result corresponding to the target image group further includes a second annotation result, which indicates the category label of the target boundary points in the bird's-eye view coordinate system. The detection result corresponding to the target image group also includes a fusion result of the second detection information and the third detection information, which indicates the category label of the boundary points of the drivable area corresponding to the target image group. For learning the drivable area boundary line category, this application can use the binary cross-entropy loss (BCE Loss) to evaluate the difference between the predicted value and the ground truth. Based on this, the binary cross-entropy loss (BCE Loss) can be determined based on the second annotation result and the fusion result. The training loss mentioned above also includes the binary cross-entropy loss (BCE Loss).

[0075] Furthermore, to address the class imbalance issue between the background and the drivable area boundary lines (obstacle boundaries, extension lines), a higher weight can be assigned to the boundary lines, and a lower weight to the background. The formula for calculating BCE Loss is as follows: ; in, These are the weights of the positive samples (boundary lines). y is the weight of the negative sample (background), y is the true value (as in the second annotation result), and p is the predicted value (as in the above fusion result in the detection result).

[0076] In this embodiment of the application, the drivable area detection network can be trained for at least one round based on the above training dataset until a preset convergence condition is met. The convergence condition may be, for example, the number of training iterations reaching a threshold, or the detection accuracy reaching a preset accuracy, etc., which are not specifically limited here.

[0077] Figure 5 This is a flowchart illustrating the method for detecting drivable areas provided in an embodiment of this application. Figure 5 As shown, the detection method includes: S501: Utilize the feature extraction network in the pre-trained drivable area detection network to extract the first feature map of each of the acquired multi-path surround view images; wherein, the drivable area detection network also includes a viewpoint transformation network and a segmentation head network; S503: Using a view transformation network, each first feature map is converted into a second feature map in the bird's-eye view space; S505: Determine the target feature map of the multi-path loop view image based on each second feature map; S507: Using a segmentation head network, drivable area detection is performed based on the target feature map, and detection results are generated. The detection results indicate the boundary points of the drivable area corresponding to the multi-path surround view image.

[0078] For an explanation of steps S501 to S507, please refer to the relevant descriptions above, and they will not be repeated here.

[0079] Figure 5 The corresponding implementation provides a solution that extracts a first feature map from each of the acquired multi-view surround-view images using a feature extraction network within a pre-trained drivable area detection network. Then, using a viewpoint transformation network within the drivable area detection network, each first feature map is transformed into a second feature map in a bird's-eye view space. Finally, based on these second feature maps, the target feature map of the multi-view surround-view images is determined. This integrates scattered multi-view image features into a unified BEV feature representation, eliminating feature differences and spatial misalignment between different viewpoints. This unified representation method allows the model to directly learn global environmental relationships from a bird's-eye view, adapting to complex spatial layouts in parking scenarios (such as narrow passages and multiple obstacles), and improving the model's adaptability to complex scenarios. Therefore, it enables more accurate environmental perception, providing a more efficient and reliable perception foundation for autonomous driving parking systems, and propelling the technology towards large-scale commercialization.

[0080] This application proposes a method for detecting and generating ground truth for drivable freespace in autonomous driving based on battery electric vehicles (BEVs). For detection, a BEV feature extraction network is constructed based on multi-channel (e.g., four-channel) surround-view fisheye images. Through hard sample mining and data augmentation strategies, the accuracy of drivable freespace detection in complex scenarios is significantly improved. For ground truth generation, a semi-automatic annotation method based on static high-precision maps is proposed. Mapping map annotations to BEV images is achieved through spatiotemporal alignment, significantly improving annotation efficiency and ground truth accuracy compared to single-frame annotation methods. The combination of these two technologies provides a complete solution for drivable freespace identification in autonomous driving.

[0081] This application innovatively addresses the pain points of related technologies by using static map projection and spatiotemporal alignment technology with multiple sensors (multi-channel surround-view cameras, LiDAR, etc.). It automatically generates BEV ground truth values ​​using static map annotation information, avoiding manual frame-by-frame annotation and significantly improving ground truth production efficiency. Ray casting and multi-view fusion ensure ground truth accuracy. A unified representation of the BEV feature space adapts to complex lighting and occlusion scenarios, enhancing algorithm robustness. The application has been implemented in engineering, and the optimized network architecture maintains high accuracy while meeting real-time requirements. This application achieves end-to-end optimization from data annotation to perception algorithms, providing a more efficient and reliable perception foundation for autonomous driving parking systems and propelling the technology towards large-scale commercialization.

[0082] It is important to note that the technical characteristics of cascaded forests and stacked ensemble learning methods inherently limit their effectiveness: traditional machine learning architectures have limited feature representation capabilities, making them ill-suited for complex and ever-changing parking scenarios; reliance on manual feature engineering leads to insufficient generalization ability; and while multi-layered cascaded structures improve model stability, they sacrifice computational efficiency. These technical characteristics collectively result in objective defects such as limited detection accuracy, significant computational latency, and insufficient scene adaptability. This application, however, overcomes the feature representation bottleneck of traditional methods through a deep learning architecture, achieving more accurate environmental perception; it replaces manual feature engineering with an end-to-end learning approach, significantly improving model generalization ability; and it optimizes network structure design, ensuring both accuracy and real-time performance. These innovative technologies comprehensively enhance the system's performance in complex parking scenarios, providing a more reliable perception solution for intelligent parking systems. The static map ground truth annotation method significantly improves annotation efficiency, and combined with multi-sensor spatiotemporal alignment technology, it greatly enhances the system's adaptability to complex scenarios such as dynamic obstacles and lighting changes, providing a reliable guarantee for mass production.

[0083] Furthermore, while existing technologies using laser point clouds for drivable area detection can cope with adverse weather conditions to some extent, they still suffer from inherent limitations such as restricted environmental adaptability, insufficient spatial resolution, weak dynamic obstacle recognition capabilities, and high system costs. In contrast, the vision-based drivable area detection method proposed in this application achieves comprehensive breakthroughs in environmental adaptability, detection accuracy, dynamic obstacle recognition, and cost-effectiveness through multi-camera fusion and deep learning BEV conversion technology. The vision-based solution not only provides higher-resolution scene understanding but also effectively overcomes the performance limitations of traditional point cloud methods in complex environments through deep learning algorithms, while significantly reducing system costs, thus offering a superior solution for the commercialization of autonomous driving systems.

[0084] The solution provided in this application achieves efficient and high-precision drivable area labeling and detection through multi-sensor spatiotemporal alignment, static map annotation reuse, and ray casting algorithms. Its core advantages are: Significantly improved annotation efficiency: Obstacle outlines only need to be annotated once on a static map to automatically generate multiple frames of BEV ground truth, which is significantly more efficient than traditional frame-by-frame annotation. Reliable annotation accuracy: Strict time synchronization, high-precision map projection, and multi-view fusion ensure centimeter-level annotation accuracy and reduce human error; Algorithm performance optimization: Multi-scale feature fusion, dual-backbone network to enhance BEV features, and three-channel segmentation head to accurately identify drivable areas, obstacle boundaries and extension lines; Strong scalability and engineering feasibility: It adapts to different sensor configurations, and the adjustable parameters are suitable for various scenarios (such as parking and urban roads), while also ensuring real-time performance.

[0085] This application embodiment also provides a truth value generation device for a drivable area, including: The first acquisition unit is configured to acquire the annotation information of a static map. The static map is constructed based on multi-channel surround view video data, LiDAR data and positioning data collected by the ground truth vehicle. The annotation information includes obstacle outlines. The second acquisition unit is configured to acquire multiple image groups, which are obtained by synchronously extracting frames from multi-channel surround view video data. Each image group includes multi-channel surround view images acquired simultaneously. The determining unit is configured to, for each of the above multiple image groups, determine the coordinate points of each obstacle contour in the vehicle coordinate system based on the pose information in the lidar data that matches the target timestamp of the target image group. The detection unit is configured to use a ray projection method in the vehicle coordinate system, taking the installation position of the multi-channel surround view camera corresponding to the multi-channel surround view video data as the observation point, and based on the coordinate points within the preset visible range, determine the drivable area under different viewing angles of the multi-channel surround view camera; take the intersection of each drivable area as the target drivable area corresponding to the target image group, determine the target boundary points of the target drivable area in the bird's-eye view coordinate system, and generate the true value results corresponding to the target image group that indicate each target boundary point.

[0086] This application embodiment also provides a drivable area detection device, including: The extraction unit is configured to extract the first feature map of each of the acquired multi-path surround view images using the feature extraction network in the pre-trained drivable region detection network; wherein the drivable region detection network also includes a viewpoint transformation network and a segmentation head network; The transformation unit is configured to use a view transformation network to transform each first feature map into a second feature map in the bird's-eye view space; The determining unit is configured to determine the target feature map of the multi-path loop view image based on each second feature map; The detection unit is configured to use a segmentation head network to detect drivable regions based on the target feature map and generate detection results. The detection results indicate the boundary points of the drivable regions corresponding to the multi-path surround view images.

[0087] This application embodiment also provides a training device for a drivable area detection network, including: The acquisition unit is configured to acquire a training dataset, wherein each training data includes an image group and the ground truth result corresponding to the image group. The image group includes multiple surround view images acquired simultaneously, and the ground truth result indicates the target boundary point of the target drivable area corresponding to the image group in the bird's-eye view coordinate system. The extraction unit is configured to extract the first feature map of each loop view image of the target image group in the training dataset using the feature extraction network in the drivable region detection network; wherein the drivable region detection network also includes a viewpoint transformation network and a segmentation head network; The transformation unit is configured to use a view transformation network to transform each first feature map into a second feature map in the bird's-eye view space; The determining unit is configured to determine the target feature map of the target image group based on each second feature map; The detection unit is configured to use a segmentation head network to detect drivable regions based on the target feature map, generate detection results, and indicate the boundary points of the drivable regions corresponding to the target image group. The adjustment unit is configured to adjust the parameters of the drivable area detection network with the goal of minimizing the difference between the ground truth and the detection results corresponding to the target image group.

[0088] For explanations of the implementation methods and beneficial effects of the various devices and units mentioned above, please refer to the relevant descriptions in the preceding text; they will not be repeated here.

[0089] This application also provides an autonomous driving system, which includes the aforementioned detection device. Further, the autonomous driving system is an autonomous parking system.

[0090] This application also provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, it implements the method of any of the above embodiments.

[0091] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, the computer program being loaded by a processor to perform the steps of any of the methods in the above embodiments.

[0092] In the embodiments of this application, the storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0093] This application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the method of any of the above embodiments.

[0094] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0095] The foregoing has provided a detailed description of the ground truth generation method, detection method, and network training method for drivable areas provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A true value generation method of a drivable area, characterized by, include: Obtain annotation information for a static map, which is constructed based on multi-channel surround-view video data, lidar data, and positioning data collected by a ground truth vehicle. The annotation information includes obstacle outlines. Multiple image groups are acquired by synchronously extracting frames from the multi-channel surround view video data, and each image group includes simultaneously acquired multi-channel surround view images. For each target image group in the plurality of image groups, based on the pose information in the lidar data that matches the target timestamp of the target image group, the coordinate points of each obstacle contour in the vehicle coordinate system are determined; In the vehicle coordinate system, using the ray projection method, with the installation position of the multi-channel surround view camera corresponding to the multi-channel surround view video data as the observation point, the drivable area under different angles of the multi-channel surround view camera is determined based on the coordinate points within the preset visible range. The intersection of each drivable region is taken as the target drivable region corresponding to the target image group. The target boundary points of the target drivable region in the bird's-eye view coordinate system are determined, and the true value results indicating each target boundary point corresponding to the target image group are generated.

2. The true value generation method according to claim 1, characterized by, After taking the intersection of each of the drivable areas as the target drivable area corresponding to the target image group, the method further includes: determining the category label corresponding to each boundary point of the target drivable area based on the obstacle contour, wherein the category label is used to characterize the obstacle boundary or extension line; The truth value results also indicate the category label corresponding to each of the target boundary points.

3. The true value generation method according to claim 1, characterized by, The acquisition of multiple image groups includes: Extract multiple frames of surround view images from one surround view video stream of the multi-channel surround view video data; For each first image in the multi-frame surround view images, based on the first timestamp of the first image, a second image is selected from each surround view video other than the first surround view video in the multi-channel surround view video data, and the first image and each second image are combined into the image group, wherein the target timestamp of the image group is the first timestamp; wherein the difference between the second timestamp of the second image and the first timestamp is less than or equal to the time difference threshold.

4. The true value generation method according to claim 1, characterized by, Pose information matching the target timestamp of the target image group includes the first pose information corresponding to the two timestamps with the smallest difference from the target timestamp in the lidar data; Based on the pose information in the lidar data that matches the timestamp of the target image group, the coordinates of each obstacle contour in the vehicle coordinate system are determined, including: Interpolate each of the first pose information to obtain the second pose information corresponding to the target timestamp; Based on the second pose information, the coordinate points of each obstacle contour in the vehicle coordinate system are determined.

5. The truth generation method of claim 1, wherein, Also includes: Filter out the coordinate points that are outside the preset visible range from each of the coordinate points; The method employs a light projection technique, using the installation position of the multi-channel surround-view camera corresponding to the multi-channel surround-view video data as the observation point. Based on the coordinate points within a preset visible range, it determines the drivable area from different angles of the multi-channel surround-view camera, including: Extract vertices from the boundary of the preset visible range and each of the coordinate points within the preset visible range; For each target camera in the multi-channel panoramic camera, the vertices are sorted according to their polar angles relative to the observation point corresponding to the target camera; Rays are emitted from the observation point to each vertex in the order described. The nearest occlusion point of each ray is detected, and the drivable area from the perspective of the target camera is determined based on the nearest occlusion point of each ray.

6. A travelable area detection method characterized by comprising: include: The first feature map of each of the acquired multi-path surround view images is extracted using the feature extraction network in the pre-trained drivable area detection network; wherein, the drivable area detection network also includes a view transformation network and a segmentation head network; Using the aforementioned perspective transformation network, each of the first feature maps is converted into a second feature map in the bird's-eye view space; The target feature map of the multi-path loop view image is determined based on each of the second feature maps; Using the segmentation head network, drivable region detection is performed based on the target feature map, and detection results are generated. The detection results indicate the boundary points of the drivable region corresponding to the multi-path surround view image.

7. The detection method according to claim 6, characterized in that, Using a pre-trained feature extraction network within a drivable region detection network, the first feature map of each of the acquired multi-path surround view images is extracted, including: The feature extraction network is used to extract multi-scale features from each loop view image, and the multi-scale features are fused into the first feature map.

8. The detection method according to claim 7, characterized in that, The feature extraction network includes a fast feature network and a feature pyramid network corresponding to the multi-path loop images, respectively. Using the feature extraction network, multi-scale features of each loop view image are extracted, and the multi-scale features are fused into the first feature map, including: The multi-scale features of each circumferential image are extracted using the fast feature network corresponding to each circumferential image, and the multi-scale features are fused into the first feature map using the feature pyramid network corresponding to each circumferential image.

9. The detection method according to claim 7, characterized in that, The segmentation head network is a three-channel segmentation head network, and the detection result also includes the category label of the boundary point, which is used to characterize the obstacle boundary or extension line.

10. A training method of a drivable area detection network, characterized by, include: Obtain a training dataset, wherein each training data includes an image group and the ground truth result corresponding to the image group, the image group includes multiple surround view images acquired simultaneously, and the ground truth result indicates the target boundary point of the target drivable area corresponding to the image group in the bird's-eye view coordinate system; The feature extraction network in the driving area detection network is used to extract the first feature map of each loop view image of the target image group in the training dataset; wherein, the driving area detection network further includes a view transformation network and a segmentation head network; Using the aforementioned perspective transformation network, each of the first feature maps is converted into a second feature map in the bird's-eye view space; The target feature map of the target image group is determined based on each of the second feature maps; Using the segmentation head network, drivable region detection is performed based on the target feature map, and detection results are generated. The detection results indicate the boundary points of the drivable region corresponding to the target image group. The parameters of the drivable area detection network are adjusted with the goal of minimizing the difference between the ground truth result and the detection result corresponding to the target image group.

Images