Image processing method, device and equipment for autonomous vehicle and medium

Abstract

The present disclosure provides an image processing method, device, equipment and medium for an autonomous vehicle, the image processing method comprising: determining an initial sample set, at least part of the vehicle surrounding scene initial sample images in the initial sample set comprising a first target obstacle, the size of the first target obstacle being smaller than a preset size; preprocessing the initial sample set to obtain a training sample set, the training sample set comprising a plurality of vehicle surrounding scene training sample images and corresponding training sample labels, the preprocessing being used to increase the number of vehicle surrounding scene training sample images comprising the first target obstacle and / or increase the number of the first target obstacles in the vehicle surrounding scene training sample images; training an initial autonomous driving obstacle recognition model based on the training sample set to obtain an autonomous driving obstacle recognition model; inputting a surrounding scene image into the autonomous driving obstacle recognition model to obtain an obstacle recognition result, so as to improve the recognition accuracy.

Description

Technical Field

[0001] This disclosure relates to the field of autonomous driving technology, and in particular to an image processing method, apparatus, device and medium for autonomous vehicles. Background Technology

[0002] In the field of autonomous driving, vehicles can collect data on their surrounding environment and status to inform subsequent decisions.

[0003] Currently, autonomous driving technology can perceive and identify the position and size of obstacles around a vehicle. It can identify large targets in images, such as cars and trucks, with relatively high accuracy, but the accuracy of identifying small targets, such as people, bicycles, and traffic cones, is lower. Summary of the Invention

[0004] To overcome the problems existing in related technologies, this disclosure provides an image processing method, apparatus, device and medium for autonomous vehicles.

[0005] According to a first aspect of this disclosure, a method for image processing for an autonomous vehicle is provided, the image processing method comprising:

[0006] An initial sample set is determined, which includes multiple initial sample images of the scene around the vehicle and corresponding initial sample labels. At least some of the initial sample images of the scene around the vehicle include a first target obstacle, the size of which is smaller than a preset size.

[0007] The initial sample set is preprocessed to obtain a training sample set, which includes multiple training sample images of vehicle surrounding scenes and corresponding training sample labels. The preprocessing is used to make the number of training sample images of vehicle surrounding scenes including the first target obstacle greater than the number of initial sample images of vehicle surrounding scenes including the first target obstacle, and / or to make the number of the first target obstacle in the training sample images of vehicle surrounding scenes greater than the number of the first target obstacle in the corresponding initial sample images of vehicle surrounding scenes.

[0008] Based on the training sample set, the initial autonomous driving obstacle recognition model is trained to obtain the autonomous driving obstacle recognition model;

[0009] Acquire images of the surrounding scene of the autonomous vehicle;

[0010] The surrounding scene image to be identified is input into the autonomous driving obstacle recognition model to obtain the obstacle recognition result.

[0011] In some embodiments of this disclosure, the preprocessing of the initial sample set includes:

[0012] The initial sample image of the scene surrounding the vehicle, including at least a portion of the first target obstacle, is oversampled to obtain a new sample image;

[0013] The newly added sample images are labeled to obtain sample labels corresponding to the newly added sample images. Each newly added sample image is determined as a training sample image of the scene around the vehicle, and the sample label corresponding to the newly added sample image is determined as the training sample label.

[0014] In some embodiments of this disclosure, each of the initial sample images of the scene surrounding the vehicle corresponds to a different time information, and the oversampling of at least a portion of the initial sample images of the scene surrounding the vehicle, including the first target obstacle, includes:

[0015] Frame interpolation is performed between the two initial sample images of the vehicle surrounding scene corresponding to two adjacent time information, so as to add at least one new sample image between the two initial sample images of the vehicle surrounding scene.

[0016] Labeling the newly added sample images includes:

[0017] Obtain the first pose information and first time information corresponding to the two initial sample images of the scene around the vehicle before and after the newly added sample image;

[0018] Determine the second pose information and second time information corresponding to the newly added sample image;

[0019] Based on the initial sample labels corresponding to the two initial sample images of the scene around the vehicle, the second pose information, and the second time information, the sample label corresponding to the newly added sample image is determined.

[0020] In some embodiments of this disclosure, the preprocessing of the initial sample set includes:

[0021] Extract the first target obstacle image corresponding to the first target obstacle from the initial sample image of the scene around the vehicle;

[0022] The first target obstacle image is merged with at least one of the other initial sample images of the scene around the vehicle in the initial sample set to obtain a merged sample image.

[0023] The merged sample image is used to identify a first target obstacle, and the sample label corresponding to the merged sample image is determined. The merged sample image is then used as a training sample image of the scene around the vehicle, and the sample label corresponding to the merged sample image is used as the training sample label.

[0024] In some embodiments of this disclosure, the first target obstacle image is merged with at least one of the other initial sample images of the vehicle surrounding scene in the initial sample set to obtain a merged sample image, including:

[0025] Obtain the first point cloud corresponding to the first target obstacle in the initial sample image of the scene around the vehicle, wherein the first point cloud is located at a first position;

[0026] Copy the first point cloud to the second position to obtain the second point cloud;

[0027] The second point cloud is projected onto the initial sample image of the scene surrounding the vehicle to obtain the projected image;

[0028] The target location is determined based on the projected image and the image of the first target obstacle corresponding to the first target obstacle;

[0029] The projected image is merged with the initial sample image of the scene around the vehicle at the target location to obtain the merged sample image.

[0030] In some embodiments of this disclosure, at least a portion of the initial sample images of the scene surrounding the vehicle include a second target obstacle, the second target obstacle being larger in size than the first target obstacle. The preprocessing of the initial sample set includes:

[0031] The camera parameters corresponding to the initial sample image of the scene around the vehicle, including at least a portion of the second target obstacle, are adjusted to reduce the second target obstacle to the size of the first target obstacle.

[0032] In some embodiments of this disclosure, the camera parameters include focal length and / or distortion parameters.

[0033] In some embodiments of this disclosure, the autonomous driving obstacle recognition model includes:

[0034] A backbone network is used to extract features from the surrounding scene image to obtain a first feature map;

[0035] A multi-scale feature extraction module is used to perform multi-scale feature extraction on the first feature map to obtain a second feature map;

[0036] An encoding network is used to extract features from the second feature map to obtain a third feature map;

[0037] The output network is used to generate the obstacle recognition result based on the third feature map.

[0038] In some embodiments of this disclosure, the backbone network includes at least one first convolutional layer and at least one spatial depth conversion module, wherein the spatial depth conversion module includes a spatial-depth conversion layer and a second convolutional layer.

[0039] In some embodiments of this disclosure, at least one of the first convolutional layer and / or the second convolutional layer is a deformable convolutional layer.

[0040] In some embodiments of this disclosure, the multi-scale feature extraction module includes a densely connected feature pyramid network.

[0041] According to a second aspect of this disclosure, an image processing apparatus for an autonomous vehicle is provided, the image processing apparatus comprising:

[0042] The determination module is configured to determine an initial sample set, which includes multiple initial sample images of the vehicle surrounding scene and corresponding initial sample labels. At least some of the initial sample images of the vehicle surrounding scene include a first target obstacle, the size of which is smaller than a preset size.

[0043] A preprocessing module is configured to preprocess the initial sample set to obtain a training sample set, the training sample set including multiple training sample images of vehicle surrounding scenes and corresponding training sample labels. The preprocessing is used to make the number of training sample images of vehicle surrounding scenes including the first target obstacle greater than the number of initial sample images of vehicle surrounding scenes including the first target obstacle, and / or make the number of the first target obstacle in the training sample images of vehicle surrounding scenes greater than the number of the first target obstacle in the corresponding initial sample images of vehicle surrounding scenes.

[0044] The training module is configured to train the initial autonomous driving obstacle recognition model based on the training sample set to obtain the autonomous driving obstacle recognition model.

[0045] The acquisition module is configured to acquire images of the surrounding scene of the autonomous vehicle;

[0046] The recognition module is configured to input the surrounding scene image into the autonomous driving obstacle recognition model to obtain obstacle recognition results.

[0047] According to a third aspect of this disclosure, an electronic device is provided, comprising:

[0048] processor;

[0049] Memory used to store the processor's executable instructions;

[0050] The processor is configured to perform the image processing method for autonomous vehicles as described in the first aspect of this disclosure.

[0051] According to a fourth aspect of this disclosure, a non-transitory computer-readable storage medium is provided, wherein when instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the image processing method for an autonomous vehicle as described in the first aspect of this disclosure.

[0052] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects: by preprocessing the initial sample set, the number of first target obstacles smaller than a preset size is increased, thereby improving the recognition ability of the trained autonomous driving obstacle recognition model for the first target obstacle and improving the accuracy of the autonomous driving obstacle recognition model in recognizing the first target obstacle.

[0053] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0054] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0055] Figure 1 This is a flowchart illustrating an image processing method for an autonomous vehicle according to an exemplary embodiment.

[0056] Figure 2 This is a flowchart illustrating an image processing method for an autonomous vehicle according to an exemplary embodiment.

[0057] Figure 3 This is a flowchart illustrating an image processing method for an autonomous vehicle according to an exemplary embodiment.

[0058] Figure 4 This is a flowchart illustrating an image processing method for an autonomous vehicle according to an exemplary embodiment.

[0059] Figure 5 This is a block diagram illustrating an image processing apparatus for an autonomous vehicle according to an exemplary embodiment.

[0060] Figure 6 This is a block diagram illustrating an electronic device according to an exemplary embodiment. Detailed Implementation

[0061] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0062] In the field of autonomous driving, vehicles can collect information about their surrounding environment and status to inform subsequent decisions. Currently, autonomous driving technology can perceive and identify the position and size of obstacles around the vehicle. It can accurately identify large targets in images, such as cars and trucks, but its accuracy is lower for small targets, such as people, bicycles, and traffic cones.

[0063] To address the aforementioned technical issues, this disclosure provides an image processing method that preprocesses an initial sample set to increase the number of first target obstacles smaller than a preset size, thereby improving the recognition capability of the trained autonomous driving obstacle recognition model for the first target obstacle and enhancing the accuracy of the autonomous driving obstacle recognition model in recognizing the first target obstacle.

[0064] To facilitate understanding, the application scenarios of the image processing method provided in this disclosure are first described. This disclosure can be applied to various fields within the field of artificial intelligence, specifically to various scenarios requiring the localization and identification of rigid bodies in the surrounding environment. A rigid body refers to an object whose shape and size remain unchanged during motion and after being subjected to forces, and whose relative positions of all points within it remain unchanged. Rigid bodies can include, for example, vehicles, roadblocks, pedestrians, or other types of rigid bodies in a road. Embodiments of this disclosure can be applied to scenarios where the position and type of targets such as roadblocks or pedestrians around a vehicle (i.e., the user's driving vehicle) need to be estimated. It is understood that the description herein is only for the convenience of understanding the application scenarios of the embodiments of this disclosure and does not exhaustively list all application scenarios of the embodiments of this disclosure. The following description uses the application of embodiments of this disclosure in the field of autonomous driving as an example.

[0065] An exemplary embodiment of this disclosure provides an image processing method for autonomous vehicles, such as... Figure 1 As shown, the image processing method for autonomous vehicles includes:

[0066] S100. Determine an initial sample set. The initial sample set includes multiple initial sample images of the scene around the vehicle and corresponding initial sample labels. At least some of the initial sample images of the scene around the vehicle include a first target obstacle. The size of the first target obstacle is smaller than a preset size.

[0067] S200. Preprocess the initial sample set to obtain a training sample set. The training sample set includes multiple training sample images of vehicle surrounding scenes and corresponding training sample labels. The preprocessing is used to make the number of training sample images of vehicle surrounding scenes including the first target obstacle greater than the number of initial sample images of vehicle surrounding scenes including the first target obstacle, and / or make the number of first target obstacles in the training sample images of vehicle surrounding scenes greater than the number of first target obstacles in the corresponding initial sample images of vehicle surrounding scenes.

[0068] S300. Based on the training sample set, train the initial autonomous driving obstacle recognition model to obtain the autonomous driving obstacle recognition model;

[0069] S400: Acquire images of the surrounding scene of the autonomous vehicle;

[0070] S500 inputs surrounding scene images into the autonomous driving obstacle recognition model to obtain obstacle recognition results.

[0071] In step S100, the initial sample image of the vehicle's surrounding scene can be an image acquired by a camera device configured on the vehicle. This camera device includes, but is not limited to, cameras, capture cards, radar, or other types of camera equipment. The camera device can be a monocular camera system, a binocular camera system, or a multi-camera system. Alternatively, the initial sample image of the vehicle's surrounding scene can be an image obtained from a database such as the Internet. This initial sample image can be a single image or a video frame. By recognizing the initial sample image of the vehicle's surrounding scene, the 2D or 3D bounding boxes in the initial sample image, along with the category corresponding to each bounding box, are used as the initial sample label for the initial sample image. Categories include, but are not limited to, vehicles, streetlights, roadblocks, road signs, guardrails, and pedestrians. The initial sample label identifies target obstacles in the initial sample image of the vehicle's surrounding scene.

[0072] Because targets closer to the camera appear larger and occupy a larger area in the image, while targets farther away appear smaller. The first target obstacle is an obstacle in the initial sample image of the scene surrounding the vehicle that is larger than or equal to a preset size. This preset size can be arbitrary, such as a preset area percentage or perimeter value, etc. A first target obstacle can also be an object far from the camera that has a smaller area percentage in the initial sample image of the scene surrounding the vehicle, such as a very distant vehicle, a distant roadblock, road sign, guardrail, or pedestrian. Some initial sample images of the scene surrounding the vehicle may include one or more first target obstacles, some may include both first target obstacles and other target obstacles, and some may not include any target obstacles, etc.

[0073] In step S200, the initial sample set is preprocessed to obtain a training sample set. The training sample set may include multiple training sample images of the vehicle's surrounding scene and their corresponding training sample labels. The category corresponding to each 2D or 3D bounding box generated in the training sample images of the vehicle's surrounding scene is used as the training sample label for the corresponding training sample image of the vehicle's surrounding scene, to identify target obstacles in the training sample images of the vehicle's surrounding scene. The preprocessing may be a method to increase the number of first target obstacles in the initial sample set.

[0074] After preprocessing the initial sample set, the number of training sample images of the vehicle's surrounding scene in the training sample set can be greater than the number of initial sample images of the vehicle's surrounding scene in the initial sample set. That is, by increasing the number of sample images including the first target obstacle, the number of training sample images of the vehicle's surrounding scene including the first target obstacle is greater than the number of initial sample images of the vehicle's surrounding scene including the first target obstacle. Alternatively, the number of training sample images of the vehicle's surrounding scene can be equal to the number of initial sample images of the vehicle's surrounding scene, by increasing the number of first target obstacles, so that the number of first target obstacles in the training sample images of the vehicle's surrounding scene is greater than the number of first target obstacles in the corresponding initial sample images of the vehicle's surrounding scene. After preprocessing, because the number of first target obstacles in the training sample set is greater than the number of first target obstacles in the initial sample set, the number of training sample labels is greater than the number of initial sample labels. By preprocessing the initial sample set and increasing the number of first target obstacles smaller than a preset size, the ability of the trained autonomous driving obstacle recognition model to recognize the first target obstacle is improved, thereby increasing the accuracy of the autonomous driving obstacle recognition model in recognizing the first target obstacle.

[0075] In step S300, the training sample set is input into the initial autonomous driving obstacle recognition model. Based on the training sample images of the vehicle's surrounding scenes and the training sample labels, the initial autonomous driving obstacle recognition model can be trained using a loss function. The training setup uses multiple training sample images of the vehicle's surrounding scenes from the training sample set and the training sample labels for each image to iteratively train the initial autonomous driving obstacle recognition model until the convergence condition of the loss function is met, thus obtaining the autonomous driving obstacle recognition model. The loss function can specifically be the L1 loss function, the cross-entropy loss function, and / or other types of loss functions. Because the number of first target obstacles smaller than a preset size increases in all the training sample labels of the vehicle's surrounding scenes, the trained autonomous driving obstacle recognition model's ability to recognize the first target obstacle is improved, thereby increasing the accuracy of the autonomous driving obstacle recognition model in recognizing the first target obstacle.

[0076] In step S400, the surrounding scene image of the autonomous vehicle can be an image acquired by image acquisition equipment configured on the vehicle. The image acquisition equipment includes, but is not limited to, cameras, acquisition cards, radar, or other types of image acquisition equipment. The image acquisition equipment can be a monocular camera system, a binocular camera system, or a multi-view camera system. The surrounding scene image can be a single image or a video frame from a video.

[0077] In step S500, the autonomous vehicle may be pre-configured with a trained autonomous driving obstacle recognition model. After acquiring images of the surrounding scene of the autonomous vehicle, these images are input into the autonomous driving obstacle recognition model. The model can use formalized information to reason based on a reasoning control strategy, such as through search and matching, to obtain obstacle recognition results. These results indicate whether the surrounding scene images contain a first target obstacle or other target obstacles, serving as a basis for subsequent decisions.

[0078] In one exemplary embodiment, such as Figure 2 As shown, in step S200, the initial sample set is preprocessed, including:

[0079] S201. Oversample the initial sample image of the scene around the vehicle, including at least part of the first target obstacle, to obtain a new sample image;

[0080] S202. Label the newly added sample images to obtain the sample labels corresponding to the newly added sample images. Determine each newly added sample image as a training sample image of the scene around the vehicle, and determine the sample label corresponding to the newly added sample image as the training sample label.

[0081] In step S201, either the initial sample images of the vehicle surrounding scene including the first target obstacle can be oversampled, or all the initial sample images of the vehicle surrounding scene including the first target obstacle can be oversampled. Oversampling refers to any technique that generates more image samples than the initial sample images of the vehicle surrounding scene including the first target obstacle, thereby obtaining new sample images. The new sample images include the first target obstacle.

[0082] In step S202, feature extraction is performed on the newly added sample images, such as generating 2D or 3D bounding boxes of the first target obstacle in the newly added sample images, and generating a category corresponding to each bounding box to obtain the sample label corresponding to the newly added sample image. Each newly added sample image is determined as a training sample image of the vehicle's surrounding scene, and the sample label corresponding to the newly added sample image is determined as the training sample label, so as to increase the number of training sample images of the vehicle's surrounding scene including the first target obstacle, making the number of training sample images of the vehicle's surrounding scene including the first target obstacle greater than the number of initial sample images of the vehicle's surrounding scene including the first target obstacle, thereby improving the recognition ability of the trained autonomous driving obstacle recognition model for the first target obstacle and improving the accuracy of the autonomous driving obstacle recognition model in recognizing the first target obstacle.

[0083] In some possible implementations, each initial sample image of the scene around the vehicle corresponds to a different time information. In step S201, oversampling is performed on at least a portion of the initial sample images of the scene around the vehicle, including the first target obstacle, including:

[0084] S201-1. Perform frame interpolation between the initial sample images of the two vehicle surrounding scenes corresponding to two adjacent time information, so as to add at least one new sample image between the two initial sample images of the vehicle surrounding scenes.

[0085] In this embodiment, each initial sample image of the scene around the vehicle is a video frame or an image captured by continuous shutter speed, so that each initial sample image of the scene around the vehicle corresponds to different time information to characterize different initial sample images of the scene around the vehicle, as well as the time relationship between two initial sample images of the scene around the vehicle.

[0086] In step S201-1, the two initial sample images of the vehicle surrounding scene corresponding to two adjacent time information may have different positions and pixels for the first target obstacle. For example, for a moving pedestrian, the position and posture in the image are different one second and the next. Frame interpolation is performed between the two initial sample images of the vehicle surrounding scene corresponding to two adjacent time information. That is, using the information of the first target obstacle in the two initial sample images of the vehicle surrounding scene corresponding to two adjacent time information, such as the position information of the 2D or 3D envelope, motion prediction and motion compensation are used to generate new video frames or images, thereby adding at least one new sample image between the two initial sample images of the vehicle surrounding scene to increase the number of sample images including the first target obstacle.

[0087] In step S202, the newly added sample images are labeled, including:

[0088] S202-1. Obtain the first pose information and first time information corresponding to the initial sample images of the scene surrounding the two vehicles before and after the newly added sample image;

[0089] S202-2, Determine the second pose information and second time information corresponding to the newly added sample image.

[0090] S202-3. Based on the initial sample labels, second pose information, and second time information corresponding to the initial sample images of the surrounding scenes of the two vehicles, determine the sample labels corresponding to the newly added sample images.

[0091] In step S202-1, since the newly added sample image is obtained by interpolating frames between the initial sample images of the vehicle's surrounding scene corresponding to two adjacent time information points, the time information corresponding to the two initial sample images of the vehicle's surrounding scene before and after the newly added sample image is obtained as the first time information. The first pose information refers to the position and pose information of the camera equipment that acquired the two initial sample images of the vehicle's surrounding scene. The pose information refers to the tilt angle of the camera equipment when capturing the initial image of the vehicle's surrounding scene. The initial sample images of the vehicle's surrounding scene are acquired by the camera equipment on the vehicle. When acquiring the initial sample images of the vehicle's surrounding scene, the position and pose of the camera equipment may be different. For example, one initial sample image of the vehicle's surrounding scene may be acquired by a monocular camera system, and another initial sample image of the vehicle's surrounding scene may be acquired by a binocular camera system, and so on.

[0092] In step S202-2, the newly added sample image is generated by frame interpolation between initial sample images of the vehicle surrounding scene corresponding to two adjacent time information, through motion prediction and motion compensation. Using the first time information and first pose information corresponding to the two initial sample images of the vehicle surrounding scene, the frame interpolation model can calculate the amount of movement of the pixels of the first target obstacle in one initial sample image to the other, or the change in the position information corresponding to the 2D or 3D envelope of the first target obstacle in the two initial sample images, to determine the position of the first target obstacle in the newly added sample image, and thus determine the second pose information and second time information corresponding to the newly added sample image.

[0093] In step S202-3, based on the initial sample labels, second pose information and second time information corresponding to the initial sample images of the scene around the two vehicles, the 2D or 3D envelope of the first target obstacle in the newly added sample image can be determined, and the category corresponding to each envelope can be generated to obtain the sample label corresponding to the newly added sample image.

[0094] In one exemplary embodiment, such as Figure 3 As shown, in step S200, the initial sample set is preprocessed, including:

[0095] S210. Extract the first target obstacle image corresponding to the first target obstacle from the initial sample image of the scene around the vehicle;

[0096] S220. Merge the first target obstacle image with at least one of the initial sample images of other vehicle surrounding scenes in the initial sample set to obtain a merged sample image.

[0097] S230. Perform first target obstacle recognition on the merged sample image, determine the sample label corresponding to the merged sample image, determine the merged sample image as the training sample image of the scene around the vehicle, and determine the sample label corresponding to the merged sample image as the training sample label.

[0098] In step S210, a feature extraction network can be used to extract a feature map within a 2D or 3D envelope that is smaller than a preset size from the feature map of the initial sample image of the scene around the vehicle, based on the position and size information of the 2D or 3D envelope output by the region proposal network (RPN). The image corresponding to this feature map is then used as the first target obstacle image and extracted.

[0099] In step S220, the image of the first target obstacle is merged with one or more initial sample images of other vehicle surrounding scenes in the initial sample set. The initial sample images of other vehicle surrounding scenes may or may not include the first target obstacle; the viewpoint of the initial sample images of other vehicle surrounding scenes may be the same as or different from the viewpoint of the initial sample image of the vehicle surrounding scene containing the first target obstacle image. The image merging method may include copying and pasting the first target obstacle image into the initial sample images of other vehicle surrounding scenes, or copying point clouds or pixels into the initial sample images of other vehicle surrounding scenes, etc., to obtain a merged sample image.

[0100] In step S230, the merged sample image is used to identify the first target obstacle. This can be achieved, for example, by using a feature extraction network. The feature extraction network can be a computationally intensive and parameter-intensive 3D obstacle detection model. The feature extraction network performs convolution operations on the merged sample image to obtain its feature map, which is then input into the RPN. Based on the feature map, the RPN outputs the position and size information of one or more 2D or 3D bounding boxes. These multiple 2D or 3D bounding boxes in the merged sample image are compared and matched with the 2D or 3D bounding boxes in the initial sample image of the vehicle's surrounding scene. This allows the identification of the added 2D or 3D bounding boxes, thus recognizing the first target obstacle in the merged sample image. The feature extraction network also generates the category corresponding to each added 2D or 3D bounding box to determine the sample label for the merged sample image. The merged sample image is determined as the training sample image of the vehicle surrounding scene, and the sample label corresponding to the merged sample image is determined as the training sample label. This makes the number of first target obstacles in the training sample image of the vehicle surrounding scene greater than the number of first target obstacles in the corresponding initial sample image of the vehicle surrounding scene. The number of first target obstacles smaller than the preset size is increased, thereby improving the recognition ability of the trained autonomous driving obstacle recognition model for the first target obstacle and improving the accuracy of the autonomous driving obstacle recognition model in recognizing the first target obstacle.

[0101] In some possible implementations, step S220 involves merging the first target obstacle image with at least one of the initial sample images of other vehicle surrounding scenes in the initial sample set to obtain a merged sample image, including:

[0102] S221. Obtain the first point cloud corresponding to the first target obstacle in the initial sample image of the scene around the vehicle. The first point cloud is located at the first position.

[0103] S222. Copy the first point cloud to the second position to obtain the second point cloud;

[0104] S223. Project the second point cloud onto the initial sample image of the scene surrounding the vehicle to obtain the projected image.

[0105] S224. Determine the target position based on the projected image and the image of the first target obstacle corresponding to the first target obstacle;

[0106] S225. The projected image at the target location is merged with the initial sample image of the scene around the vehicle to obtain the merged sample image.

[0107] In step S221, a point cloud acquisition device can be used to extract the first point cloud corresponding to the first target obstacle in the initial sample image of the scene around the vehicle. The point cloud acquisition device can be a LiDAR. Specifically, the first point cloud can be a massive set of points representing the spatial distribution and surface characteristics of the first target obstacle in the same spatial reference frame. After obtaining the spatial coordinates of each sampling point on the surface of the first target obstacle, a set of points is obtained, called a "point cloud". Alternatively, the point cloud image of the initial sample image of the scene around the vehicle can be obtained using the point cloud image function of the camera device, thereby obtaining the first point cloud corresponding to the first target obstacle. The point cloud can include three-dimensional coordinates (X, Y, Z). For example, the pixel coordinates of the first target obstacle in the initial sample image of the scene around the vehicle can be converted into first point cloud coordinates, and the first point cloud coordinates are defined as the first position.

[0108] In step S222, the first point cloud is copied to the second position. The second position can be any one or more positions in the initial sample image of the scene around the vehicle, except for the first position, or it can be a three-dimensional position outside the initial sample image of the scene around the vehicle, so that the second point cloud can be projected onto the initial sample image of the scene around the vehicle in subsequent steps.

[0109] In step S223, the second point cloud is projected onto the initial sample image of the scene surrounding the vehicle to obtain a projected image. The initial sample image of the scene surrounding the vehicle can be an image containing the first target obstacle or an image not containing the first target obstacle. The projected image is the image corresponding to the second point cloud. For example, the projected image can be an image with the same features as the first target obstacle image but a different spatial location.

[0110] In step S224, the target position is determined based on the projected image and the first target obstacle image corresponding to the first target obstacle. The target position is the position of the projected image in the initial sample image of the scene around the vehicle. The target position can be any one or more positions in the initial sample image of the scene around the vehicle, except for the first position.

[0111] In step S225, the target position of the projected image in the initial sample image of the vehicle surrounding scene is merged with the initial sample image of the vehicle surrounding scene, for example, by merging through overlay interpolation, to obtain a merged sample image. This ensures that the number of first target obstacles in the training sample image of the vehicle surrounding scene is greater than the number of first target obstacles in the corresponding initial sample image of the vehicle surrounding scene, increasing the number of first target obstacles smaller than a preset size. This improves the recognition ability of the trained autonomous driving obstacle recognition model for the first target obstacles, thereby enhancing the accuracy of the autonomous driving obstacle recognition model in recognizing the first target obstacles.

[0112] In one exemplary embodiment, at least a portion of the initial sample image of the scene surrounding the vehicle includes a second target obstacle, the size of which is larger than the size of the first target obstacle, such as... Figure 4 As shown, in step S200, the initial sample set is preprocessed, including:

[0113] S200-1. Adjust the camera parameters corresponding to the initial sample image of the scene around the vehicle, including at least a portion of the second target obstacle, to reduce the second target obstacle to the first target obstacle.

[0114] In this embodiment, the second target obstacle is an obstacle with a preset size or larger. For example, it is an object that is close to the camera device and whose area proportion in the initial sample image of the scene around the vehicle is greater than the preset proportion, such as a nearby vehicle, a very close roadblock, road sign, guardrail, pedestrian, etc. Some initial sample images of the scene around the vehicle may include one or more second target obstacles, and some initial sample images of the scene around the vehicle may also include a second target obstacle and a first target obstacle, etc.

[0115] In this embodiment, a virtual camera can be constructed. This virtual camera is a non-existent camera that is arranged in three-dimensional space using multiple initial sample images of the vehicle's surrounding scene, including the second target obstacle. The virtual camera can reproduce the initial sample images of the vehicle's surrounding scene from a virtual viewpoint. Since the closer the camera is to the object when capturing images, the higher the image resolution and the larger the target size, the more effective the virtual camera can be. By adjusting the parameters of the virtual camera corresponding to the initial sample images of the vehicle's surrounding scene including the second target obstacle, the size of the second target obstacle can be changed. This causes distortion in the image of the second target obstacle, thereby reducing the second target obstacle to the size of the first target obstacle. For example, the camera parameters can include the focal length and / or distortion parameters of the virtual camera. For instance, a virtual camera with a small focal length and large distortion will significantly reduce the size of edge objects in the image, resulting in a greater number of first target obstacles in the training sample images of the vehicle's surrounding scene than in the corresponding initial sample images. This increases the number of first target obstacles smaller than a preset size, thereby improving the recognition ability of the trained autonomous driving obstacle recognition model for the first target obstacle and increasing the accuracy of the autonomous driving obstacle recognition model in recognizing the first target obstacle.

[0116] In one exemplary embodiment, the autonomous driving obstacle recognition model includes:

[0117] The backbone network is used to extract features from the surrounding scene image to obtain the first feature map;

[0118] The multi-scale feature extraction module is used to extract multi-scale features from the first feature map to obtain the second feature map.

[0119] An encoding network is used to extract features from the second feature map to obtain the third feature map;

[0120] The output network is used to generate the obstacle recognition result based on the third feature map.

[0121] In this embodiment, the backbone network may include neural network models such as backbone, regnet, and resnet, which are mainly used to extract features from the surrounding scene images of the autonomous vehicle to obtain a first feature map for use by subsequent networks.

[0122] For example, the backbone network can be constructed using a spatial depth transformation module, which includes at least one first convolutional layer and at least one spatial depth transformation module, the spatial depth transformation module including a spatial-depth transformation layer and a second convolutional layer.

[0123] Each convolutional layer consists of several convolutional units, and the parameters of each unit are optimized using the backpropagation algorithm. Convolutional layers are used to extract different features from the input; through iterations of different convolutional layers, more complex features can be extracted from low-level features. The spatial-depth transformation module performs spatial and depth transformations on the surrounding scene image, so that with each pass through the spatial-depth transformation module, the number of pixels in the first feature map of the surrounding scene image is halved, thereby improving the computational speed of the backbone network. The spatial-depth transformation layer is used to enlarge, reduce, or flip the surrounding scene image.

[0124] In some possible implementations, at least one first convolutional layer and / or one second convolutional layer is a deformable convolutional layer. A deformable convolutional layer refers to a convolutional kernel with an additional direction parameter added to each element, allowing the kernel to transform into any shape, thus expanding the range of features acquired. In this embodiment, by setting at least one first convolutional layer or at least one second convolutional layer as a deformable convolutional layer, the convolutional kernel can adjust its shape according to actual target obstacles in the surrounding scene image, better extracting input features and improving the convolutional layer's adaptability and generalization ability.

[0125] The multi-scale feature extraction module is used to extract features from the first feature map at multiple scales. In this embodiment, the lower network layers have smaller receptive fields and acquire low-level feature information, resulting in better extraction of the first target obstacle. The higher network layers acquire deep feature information, leading to more accurate detection of the second target obstacle. Dilated convolutions with different dilation rates can acquire multi-scale global contextual feature information to compensate for the shortcomings of single feature detection. Different scale target detectors are designed for different output layers to complete feature extraction at multiple scales. The multi-scale feature extraction module also includes a multi-level feature fusion module, which can fuse shallow feature information, deep feature information, and global contextual feature information using their distribution characteristics to obtain the second feature map.

[0126] In some possible implementations, the multi-scale feature extraction module includes a densely connected feature pyramid network. In this embodiment, the feature pyramid network scales the first feature map at different scales to obtain an image pyramid formed by multiple layers of images. Then, features at different scales are extracted from each layer of images to obtain a second feature map. The pyramid of the first feature map is a set of images arranged in a pyramid shape with progressively decreasing resolution, all originating from the same original image, which is the first feature map. The image pyramid is obtained through stepwise downsampling until a certain termination condition is reached. The higher the image level, the smaller the image and the lower the resolution. This ensures that features at different scales can contain rich feature information with high accuracy, resulting in a high-precision second feature map. The connections in the densely connected feature pyramid network are densely connected to enhance the interaction and fusion of features at different scales.

[0127] The encoding network is used to extract features from the second feature map to obtain the third feature map. The encoding network can be, for example, a Transformer model. The encoding network can perform weight calculation and weighted summation on the feature information in the second feature map. It can also use local sparse attention to reduce the complexity of the image recognition model and improve the computation speed.

[0128] The output network is used to generate image recognition results based on the third feature map. The output network can be, for example, a head network. It makes predictions based on the third feature map, for example, by using formal information to reason according to the inference control strategy, such as by searching and matching, to obtain the obstacle recognition results output by the autonomous driving obstacle recognition model. The obstacle recognition results are used to indicate whether the surrounding scene image of the autonomous vehicle contains a first target obstacle or a second target obstacle, as a basis for subsequent decisions.

[0129] In one exemplary embodiment, such as Figure 5 As shown, an exemplary embodiment of this disclosure provides an image processing apparatus for an autonomous vehicle, the image processing apparatus being configured to perform the above-described image processing method for an autonomous vehicle, the image processing apparatus comprising:

[0130] The determination module 100 is configured to determine an initial sample set, which includes multiple initial sample images of the scene around the vehicle and corresponding initial sample labels. At least some of the initial sample images of the scene around the vehicle include a first target obstacle, the size of which is smaller than a preset size.

[0131] The preprocessing module 200 is configured to preprocess the initial sample set to obtain a training sample set, which includes multiple training sample images of vehicle surrounding scenes and corresponding training sample labels. The preprocessing is used to make the number of training sample images of vehicle surrounding scenes including the first target obstacle greater than the number of initial sample images of vehicle surrounding scenes including the first target obstacle, and / or make the number of first target obstacles in the training sample images of vehicle surrounding scenes greater than the number of first target obstacles in the corresponding initial sample images of vehicle surrounding scenes.

[0132] Training module 300 is configured to train the initial autonomous driving obstacle recognition model based on the training sample set to obtain the autonomous driving obstacle recognition model;

[0133] The acquisition module 400 is configured to acquire images of the surrounding scene of the autonomous vehicle;

[0134] The recognition module 500 is configured to input surrounding scene images into the autonomous driving obstacle recognition model to obtain obstacle recognition results.

[0135] like Figure 6 As shown, an exemplary embodiment of this disclosure provides an electronic device, such as a mobile phone, a laptop computer, a tablet computer, an autonomous vehicle, and a wearable device.

[0136] refer to Figure 6 As shown, the electronic device 600 may include one or more of the following components: a processing component 602, a memory 604, a power supply component 606, a multimedia component 608, an audio component 610, an input / output (I / O) interface 612, a sensor component 614, and a communication component 616.

[0137] Processing component 602 typically controls the overall operation of electronic device 600, such as operations associated with display, telephone calls, data communication, camera operation, and recording operations. Processing component 602 may include one or more processors 620 to execute instructions to perform all or part of the steps of the methods described above. Furthermore, processing component 602 may include one or more modules to facilitate interaction between processing component 602 and other components. For example, processing component 602 may include a multimedia module to facilitate interaction between multimedia component 608 and processing component 602.

[0138] Memory 604 is configured to store various types of data to support the operation of electronic device 600. Examples of this data include instructions for any application or method operating on electronic device 600, contact data, phonebook data, messages, pictures, videos, etc. Memory 604 can be implemented by any type of volatile or non-volatile storage terminal or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0139] Power supply component 606 provides power to various components of electronic device 600. Power supply component 606 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to electronic device 600.

[0140] Multimedia component 608 includes a screen that provides an output interface between electronic device 600 and user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors may sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation. In some embodiments, multimedia component 608 includes a front-facing camera module and / or a rear-facing camera module. When electronic device 600 is in an operating mode, such as shooting mode or video mode, the front-facing camera module and / or rear-facing camera module may receive external multimedia data. Each front-facing camera module and rear-facing camera module may be a fixed optical lens system or have focal length and optical zoom capabilities.

[0141] Audio component 610 is configured to output and / or input audio signals. For example, audio component 610 includes a microphone (MIC) configured to receive external audio signals when electronic device 600 is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals may be further stored in memory 604 or transmitted via communication component 616. In some embodiments, audio component 610 also includes a speaker for outputting audio signals.

[0142] I / O interface 612 provides an interface between processing component 602 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include, but are not limited to, home buttons, volume buttons, power buttons, and lock buttons.

[0143] Sensor assembly 614 includes one or more sensors for providing state assessments of various aspects of electronic device 600. For example, sensor assembly 614 may detect the on / off state of electronic device 600, the relative positioning of components such as the display and keypad of electronic device 600, changes in position of electronic device 600 or a component of electronic device 600, the presence or absence of user contact with electronic device 600, orientation or acceleration / deceleration of electronic device 600, and temperature changes of electronic device 600. Sensor assembly 614 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. Sensor assembly 614 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, sensor assembly 614 may also include an accelerometer, gyroscope, magnetometer, pressure sensor, or temperature sensor.

[0144] Communication component 616 is configured to facilitate wired or wireless communication between electronic device 600 and other terminals. Electronic device 600 can access wireless networks based on communication standards, such as WiFi, 2G, 3G, 4G, 6G, or combinations thereof. In one exemplary embodiment, communication component 616 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 616 also includes a near-field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0145] For example, the electronic device 600 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing terminals (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the methods described above.

[0146] In one exemplary embodiment, this disclosure also provides a non-transitory computer-readable storage medium including instructions, such as a memory 604 including instructions, which can be executed by a processor 620 of an electronic device 600 to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage terminal, etc. When the instructions in the storage medium are executed by the processor of the terminal, the terminal is able to perform the method shown in the above embodiments.

[0147] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0148] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. An image processing method for autonomous vehicles, characterized in that, The image processing method includes: An initial sample set is determined, which includes multiple initial sample images of the scene around the vehicle and corresponding initial sample labels. At least some of the initial sample images of the scene around the vehicle include a first target obstacle, the size of which is smaller than a preset size. The initial sample set is preprocessed to obtain a training sample set, which includes multiple training sample images of vehicle surrounding scenes and corresponding training sample labels. The preprocessing is used to make the number of training sample images of vehicle surrounding scenes including the first target obstacle greater than the number of initial sample images of vehicle surrounding scenes including the first target obstacle, and / or to make the number of the first target obstacle in the training sample images of vehicle surrounding scenes greater than the number of the first target obstacle in the corresponding initial sample images of vehicle surrounding scenes. Based on the training sample set, the initial autonomous driving obstacle recognition model is trained to obtain the autonomous driving obstacle recognition model; and the surrounding scene images of the autonomous vehicle are acquired. The surrounding scene image is input into the autonomous driving obstacle recognition model to obtain the obstacle recognition result; The preprocessing of the initial sample set includes: Oversampling at least a portion of the initial sample images of the vehicle surrounding scene, including a first target obstacle, to obtain new sample images includes: acquiring first pose information and first time information corresponding to two consecutive initial sample images of the vehicle surrounding scene; performing frame interpolation processing between two initial sample images of the vehicle surrounding scene corresponding to two adjacent time information based on the first pose information and the first time information, so as to add at least one new sample image between the two initial sample images of the vehicle surrounding scene; wherein the new sample image is generated by motion prediction and motion compensation, including: calculating, through a frame interpolation model, the amount of movement of a pixel of the first target obstacle in one initial sample image of the vehicle surrounding scene to another initial sample image of the vehicle surrounding scene, or the change value of the position information corresponding to the envelope of the first target obstacle in the two initial sample images of the vehicle surrounding scene, to determine the position of the first target obstacle in the new sample image, and then determining the second pose information and second time information corresponding to the new sample image; The newly added sample image is labeled to obtain the sample label corresponding to the newly added sample image, which includes: determining the sample label corresponding to the newly added sample image based on the initial sample labels corresponding to the two initial sample images of the vehicle surrounding scene before and after, the second pose information, and the second time information; Each newly added sample image is determined as a training sample image of the scene surrounding the vehicle, and the sample label corresponding to the newly added sample image is determined as the training sample label.

2. The image processing method for autonomous vehicles according to claim 1, characterized in that, In the initial sample images of the scene around the vehicle, each initial sample image of the scene around the vehicle corresponds to a different time information.

3. The image processing method for autonomous vehicles according to claim 1, characterized in that, The preprocessing of the initial sample set includes: Extract the first target obstacle image corresponding to the first target obstacle from the initial sample image of the scene around the vehicle; The first target obstacle image is merged with at least one of the other initial sample images of the scene around the vehicle in the initial sample set to obtain a merged sample image. The merged sample image is used to identify a first target obstacle, and the sample label corresponding to the merged sample image is determined. The merged sample image is then used as a training sample image of the scene around the vehicle, and the sample label corresponding to the merged sample image is used as the training sample label.

4. The image processing method for autonomous vehicles according to claim 2, characterized in that, The first target obstacle image is merged with at least one of the other initial sample images of the scene around the vehicle in the initial sample set to obtain a merged sample image, including: Obtain the first point cloud corresponding to the first target obstacle in the initial sample image of the scene around the vehicle, wherein the first point cloud is located at a first position; Copy the first point cloud to the second position to obtain the second point cloud; The second point cloud is projected onto the initial sample image of the scene surrounding the vehicle to obtain the projected image; The target location is determined based on the projected image and the image of the first target obstacle corresponding to the first target obstacle; The projected image is merged with the initial sample image of the scene around the vehicle at the target location to obtain the merged sample image.

5. The image processing method for autonomous vehicles according to claim 1, characterized in that, At least a portion of the initial sample images of the scene surrounding the vehicle include a second target obstacle, the second target obstacle being larger in size than the first target obstacle. The preprocessing of the initial sample set includes: The camera parameters corresponding to the initial sample image of the scene around the vehicle, including at least a portion of the second target obstacle, are adjusted to reduce the second target obstacle to the size of the first target obstacle.

6. The image processing method for autonomous vehicles according to claim 5, characterized in that, The camera parameters include focal length and / or distortion parameters.

7. The image processing method for autonomous vehicles according to any one of claims 1 to 6, characterized in that, The autonomous driving obstacle recognition model includes: A backbone network is used to extract features from the surrounding scene image to obtain a first feature map; A multi-scale feature extraction module is used to perform multi-scale feature extraction on the first feature map to obtain a second feature map; an encoding network is used to extract features from the second feature map to obtain a third feature map; The output network is used to generate the obstacle recognition result based on the third feature map.

8. The image processing method for autonomous vehicles according to claim 7, characterized in that, The backbone network includes at least one first convolutional layer and at least one spatial depth transformation module, wherein the spatial depth transformation module includes a spatial-depth transformation layer and a second convolutional layer.

9. The image processing method for autonomous vehicles according to claim 8, characterized in that, At least one of the first convolutional layer and / or the second convolutional layer is a deformable convolutional layer.

10. The image processing method for autonomous vehicles according to claim 7, characterized in that, The multi-scale feature extraction module includes a densely connected feature pyramid network.

11. An image processing apparatus for an autonomous vehicle, characterized in that, The image processing device includes: The determination module is configured to determine an initial sample set, which includes multiple initial sample images of the vehicle surrounding scene and corresponding initial sample labels. At least some of the initial sample images of the vehicle surrounding scene include a first target obstacle, the size of which is smaller than a preset size. A preprocessing module is configured to preprocess the initial sample set to obtain a training sample set, the training sample set including multiple training sample images of vehicle surrounding scenes and corresponding training sample labels. The preprocessing is used to make the number of training sample images of vehicle surrounding scenes including the first target obstacle greater than the number of initial sample images of vehicle surrounding scenes including the first target obstacle, and / or make the number of the first target obstacle in the training sample images of vehicle surrounding scenes greater than the number of the first target obstacle in the corresponding initial sample images of vehicle surrounding scenes. The training module is configured to train the initial autonomous driving obstacle recognition model based on the training sample set to obtain the autonomous driving obstacle recognition model. The acquisition module is configured to acquire images of the surrounding scene of the autonomous vehicle; The recognition module is configured to input the surrounding scene image into the autonomous driving obstacle recognition model to obtain obstacle recognition results; The preprocessing module is configured to oversample at least a portion of the initial sample images of the vehicle surrounding scene, including the first target obstacle, to obtain new sample images. This includes: acquiring first pose information and first time information corresponding to two consecutive initial sample images of the vehicle surrounding scene; performing frame interpolation processing between two initial sample images of the vehicle surrounding scene corresponding to two adjacent time information based on the first pose information and the first time information, to add at least one new sample image between the two initial sample images of the vehicle surrounding scene. The new sample image is generated through motion estimation and motion compensation. This includes: calculating, through a frame interpolation model, the amount of movement of a pixel of the first target obstacle in one initial sample image of the vehicle surrounding scene to another initial sample image of the vehicle surrounding scene, or the change in position information corresponding to the envelope of the first target obstacle in the two initial sample images of the vehicle surrounding scene, to determine the position of the first target obstacle in the new sample image, and then determining the second pose information and second time information corresponding to the new sample image. The newly added sample image is labeled to obtain the sample label corresponding to the newly added sample image, which includes: determining the sample label corresponding to the newly added sample image based on the initial sample labels corresponding to the two initial sample images of the vehicle surrounding scene before and after, the second pose information, and the second time information; Each newly added sample image is determined as a training sample image of the scene surrounding the vehicle, and the sample label corresponding to the newly added sample image is determined as the training sample label.

12. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to perform the image processing method for an autonomous vehicle as described in any one of claims 1 to 10.

13. A non-transitory computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device is able to perform the image processing method for an autonomous vehicle as described in any one of claims 1 to 10.

Images