Scene type identification method and device for automatic driving and computer readable storage medium

Abstract

The invention discloses an automatic driving scene type identification method and device and a computer readable storage medium, and the method comprises the steps: obtaining the map labeling data of a scene based on the positioning of a vehicle when the vehicle enters a certain scene, and obtaining the scene type according to the map labeling data; if the map labeling data of the scene is not obtained, image data sensed by the scene are obtained and recognized, and a recognized scene type is obtained; if the confidence coefficient of image recognition does not meet a preset requirement, obtaining depth information of pixel points of a two-dimensional image from the image data through a pre-trained BEV depth prediction model, forming depth distribution of the scene, and converting the depth distribution into BEV space to obtain look-around BEV data of the scene; and inputting the look-around BEV data into a pre-trained scene type identification model to obtain an identified scene type. According to the invention, the scene type identification effect is improved through the serial mode of three mechanisms.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent driving technology, and specifically relates to a scene type recognition method, device and computer-readable storage medium for automatic driving. Background Art

[0002] In order to achieve a wider range of automotive robot functions, it is necessary to design and develop cognitive modules in a targeted manner, of which scene understanding is a part of cognition. Scene understanding is responsible for identifying which type of artificially defined scene the vehicle is in based on the perception, map, positioning environment and vehicle data, and supporting the decision-making and planning module to perform corresponding decision-making and planning control tasks for specific scenarios.

[0003] The first stage function of scene understanding requires the recognition of scene types, which include venue types (such as toll booths or parking lots) and task stage types (such as determining whether the vehicle has reached the lane entrance selection stage or the barrier passing stage in the toll booth scene). Different scenes have different venues, different venues have different task stages, and different task stages have different states and tasks. The second stage function of scene understanding is to understand the scene in which the vehicle is located, as well as the tasks and execution processes required in the scene; the third stage function is to check the completion of the vehicle's tasks and determine whether to execute the next task; the fourth stage function is to update the cognitive scene after the task is completed to complete the scene switch, and send instructions to the decision-making planning module to complete the function switch, such as switching from the urban road scene function to the highway scene function.

[0004] At present, there is still no clear and open mature technical solution for scene understanding methods used in the "three-domain integration" of highway, urban and parking domains and the development of automotive robot functions. The known L2++ parking scene understanding solutions use positioning and map information for scene understanding, which is not generalizable and is only applicable to pre-set scenes. The scope of use is limited and it is impossible to understand scenes without maps, light maps and pre-set information. Summary of the invention

[0005] The technical problem to be solved by the embodiments of the present invention is to provide a method, device and computer-readable storage medium for scene type recognition in autonomous driving, so as to make full use of the perception capability, improve the recognition effect of scene types in the absence of maps or light maps, and lay a foundation for the decision-making and planning module to better complete tasks in specific scenarios and improve point-to-point autonomous driving capabilities.

[0006] In order to solve the above technical problems, the present invention provides a method for identifying scene types of autonomous driving, comprising:

[0007] When the vehicle enters a certain scene, the map annotation data of the scene is obtained based on the vehicle positioning, and the scene type is obtained according to the map annotation data;

[0008] If the map annotation data of the scene is not obtained, image data perceived by the scene is obtained and recognized to obtain the recognized scene type;

[0009] If the confidence of image recognition does not meet the preset requirements, the image data is passed through a pre-trained BEV depth prediction model to obtain the depth information of the two-dimensional image pixels, and the depth distribution of the scene is formed, and then the depth distribution is converted into the BEV space to obtain the surround view BEV data of the scene; the surround view BEV data is input into a pre-trained scene type recognition model to obtain the recognized scene type.

[0010] It can be seen from the above steps that the embodiment of the present invention uses three mechanisms of map scene annotation, image recognition, and BEV perception to realize scene type recognition in two cases: with map scene annotation information and without map or light map scene annotation information, thereby improving the reliability of scene recognition; at the same time, the scene understanding function is realized with the minimum computing power possible through the three mechanisms in series, thereby ensuring to the greatest extent that the scene understanding function can generate scene attribute results.

[0011] Preferably, when training the BEV depth prediction model, pixel points whose depth information is supplemented by a square root 3 subdivision surface algorithm and pixel points having depth information of a LiDAR point cloud are used as training samples.

[0012] Preferably, the method of supplementing the depth information by using a square root 3 subdivision surface algorithm specifically includes:

[0013] Connect any LiDAR point in the acquired LiDAR point cloud data to its two nearest LiDAR points to form a polyhedron composed of multiple triangular faces;

[0014] Subdividing the triangular surface of the polyhedron multiple times using a square root 3 subdivision surface algorithm to obtain an approximate surface mesh representing the three-dimensional contours of all pixel points of the two-dimensional image;

[0015] Depth information of each pixel is calculated based on the approximate surface mesh.

[0016] Preferably, the triangular surface of the polyhedron is subdivided multiple times using a square root 3 subdivision surface algorithm, specifically including: each time the subdivision is performed, a new vertex is inserted into each triangle, the new vertex is used as the center of gravity of the original triangle, the new vertex is connected to the three vertices of the original triangle, and three new triangles are obtained, and then the three sides of the original triangle are removed.

[0017] Preferably, calculating the depth information of each pixel point based on the approximate surface mesh specifically includes: determining the triangle corresponding to each pixel point in the approximate surface mesh, using the vertex information of the triangle and the position of the pixel point in the triangle, and estimating the depth information of the pixel point through an interpolation method.

[0018] The present invention also provides a scene type recognition device for automatic driving, comprising:

[0019] A map scene annotation recognition module is used to obtain map annotation data of a scene based on the vehicle positioning when the vehicle enters a scene, and obtain the scene type according to the map annotation data;

[0020] An image recognition module, for obtaining image data perceived for the scene and performing recognition to obtain a recognized scene type if map annotation data for the scene is not obtained;

[0021] The BEV recognition module is used to obtain the depth information of the two-dimensional image pixels through the pre-trained BEV depth prediction model if the confidence of image recognition does not meet the preset requirements, and form the depth distribution of the scene, and then convert the depth distribution into the BEV space to obtain the surround view BEV data of the scene; input the surround view BEV data into the pre-trained scene type recognition model to obtain the recognized scene type.

[0022] Preferably, when training the BEV depth prediction model, pixel points whose depth information is supplemented by a square root 3 subdivision surface algorithm and pixel points having depth information of a LiDAR point cloud are used as training samples.

[0023] Preferably, the method of supplementing the depth information by using a square root 3 subdivision surface algorithm specifically includes:

[0024] Connect any LiDAR point in the acquired LiDAR point cloud data to its two nearest LiDAR points to form a polyhedron composed of multiple triangular faces;

[0025] Subdividing the triangular surface of the polyhedron multiple times using a square root 3 subdivision surface algorithm to obtain an approximate surface mesh representing the three-dimensional contours of all pixel points of the two-dimensional image;

[0026] Depth information of each pixel is calculated based on the approximate surface mesh.

[0027] Preferably, the triangular surface of the polyhedron is subdivided multiple times using a square root 3 subdivision surface algorithm, specifically including: each time the subdivision is performed, a new vertex is inserted into each triangle, the new vertex is used as the center of gravity of the original triangle, the new vertex is connected to the three vertices of the original triangle, and three new triangles are obtained, and then the three sides of the original triangle are removed.

[0028] The present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored computer program; wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the scene type recognition method for autonomous driving.

[0029] The implementation of the present invention has the following beneficial effects: using the three mechanisms of map scene annotation, image recognition, and BEV perception to realize scene type recognition in two cases where there is map scene annotation information and where there is no map or light map scene annotation information, thereby improving the reliability of scene recognition; at the same time, the scene understanding function is realized with the minimum computing power possible through the three mechanisms in a serial manner, ensuring that the scene understanding function can generate scene attribute results to the greatest extent. The BEV perception method is improved, and the pixel depth information of the two-dimensional image is supplemented by the square root 3 subdivision surface algorithm, so that the final BEV depth distribution prediction result has a higher degree of restoration. The present invention can provide scene category information for the decision-making planning module, so that the decision-making planning module can better complete tasks in specific scenarios, improve point-to-point autonomous driving capabilities, and provide the prerequisites and cognitive basis for the "three-domain integration" of highway, city, and parking scenarios, thereby improving the user's riding experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0031] Figure 1 It is a flowchart of a method for identifying scene types for autonomous driving according to an embodiment of the present invention.

[0032] Figure 2 It is a specific flow chart of a method for identifying scene types for autonomous driving according to Embodiment 1 of the present invention. DETAILED DESCRIPTION

[0033] The following descriptions of the embodiments refer to the accompanying drawings to illustrate specific embodiments in which the present invention may be implemented.

[0034] Please refer to Figure 1As shown, the first embodiment of the present invention provides a method for identifying scene types of autonomous driving, including:

[0035] When the vehicle enters a certain scene, the map annotation data of the scene is obtained based on the vehicle positioning, and the scene type is obtained according to the map annotation data;

[0036] If the map annotation data of the scene is not obtained, image data perceived by the scene is obtained and recognized to obtain the recognized scene type;

[0037] If the confidence of image recognition does not meet the preset requirements, the image data is passed through a pre-trained BEV depth prediction model to obtain the depth information of the two-dimensional image pixels, and the depth distribution of the scene is formed, and then the depth distribution is converted into the BEV space to obtain the surround view BEV data of the scene; the surround view BEV data is input into a pre-trained scene type recognition model to obtain the recognized scene type.

[0038] From the above steps, it can be seen that the embodiment of the present invention adopts a triple serial mechanism of scene map annotation - image recognition - BEV recognition to realize scene type recognition in two cases: with map scene annotation information and without map or light map scene annotation information, so as to realize the scene understanding function with the smallest possible computing power.

[0039] It is understandable that the elements that determine a scene type are mainly static, such as roads, lane lines, traffic signs, road facilities (such as highway toll stations and parking lots) and other static objects, and these static scene features are strongly related to maps. Therefore, scene recognition mainly relies on the use of static data, and the most important static data is map data. By annotating the semantic information of the site on the map, the type of scene where the vehicle is located can be directly obtained, such as intersections, highway toll stations, etc. After completing the definition and identification of the scene, the definition and identification of the task stage are completed in the scene.

[0040] Please combine Figure 2 As shown, in the embodiment of the present invention, the definition of the scene has been completed in advance according to the needs, for example: 1. Intersections (crossroads, T-intersections, Y-intersections, U-intersections (with directions), traffic light intersections); 2. Expressway toll stations, expressways; 3. Parking lots; 4. Roundabouts; 5. Merging sections (non-Y-intersections); 6. Diverting sections (non-Y-intersections); 7. Overpasses, 8. Entry and exit ramps.

[0041] Then annotate the scenes in the map. The specific implementation method is: in the map in Opendrive or ApolloOpendrive format, cover and select the corresponding scene with a convex polygon, generate the corresponding scene data, and add the corresponding scene annotation to form the map annotation data; add the map annotation data reading function in the map parser of the autonomous driving system. When the vehicle enters a certain scene, the decision-making planning module reads the map annotation data based on the vehicle positioning. If there is map annotation data for the scene, the scene type can be obtained, the scene information can be updated, and the scene understanding is completed.

[0042] It should be noted that the scene map annotation in the embodiment of the present invention is not used for positioning, but for scene understanding, that is, understanding the scene in which the vehicle is located in order to perform the decision-making and planning tasks that need to be completed in the scene. It is a prerequisite for decision-making and planning and belongs to the cognitive link of the decision-making and planning module.

[0043] If there is no map annotation data for the scene, the next mechanism is required - image recognition. Specifically, the two-dimensional image data of the scene is obtained using the on-board camera, and the image recognition algorithm is used to identify two-dimensional static objects from the city road, such as highway toll stations (generally, toll stations are marked on the map), parking lots (if there is no parking lot information on the map, it is necessary to identify parking lots or parking areas), and other places. Then, the target detection algorithm is used to identify key objects in the venue, such as identifying warning columns and their size and position at the entrance of the toll station for self-vehicle obstacle avoidance. After identifying the scene type represented by the current image data, if the confidence of the image recognition meets the corresponding requirements, the scene type is obtained, the scene information is updated, and the scene understanding is completed.

[0044] If the confidence of image recognition does not meet the preset requirements, the embodiment of the present invention provides a third mechanism - identifying the scene type through the perception BEV method, that is, mapping the two-dimensional image data obtained by the vehicle camera to the three-dimensional BEV space, obtaining the three-dimensional information of the scene in the BEV space with a more comprehensive perspective to identify the scene type, and identify the stage at which the current vehicle has reached in executing the task. Specifically, a Transformer neural network is used to obtain the depth information of the two-dimensional image pixels through a depth prediction algorithm (Lift-Splat-Shoot, LSS) from the original two-dimensional image data, and form the depth distribution of the scene where the vehicle is located, and then convert it to the two-dimensional BEV space to obtain the surround view BEV data of the scene.

[0045] Among them, in the process of generating the depth distribution, the LiDAR point cloud is used as the true value of the image pixel depth for supervised training. Considering that the LiDAR point cloud is discrete and relatively sparse, there are still a large number of pixels that lack depth information, which is not conducive to the training of the depth prediction model. For this reason, it is necessary to supplement the depth information of some pixels without LiDAR point clouds with "pseudo point clouds" to establish a denser pixel depth distribution with more depth information, so as to obtain a more realistic pixel depth distribution map. Therefore, the embodiment of the present invention uses the square root 3 subdivision surface algorithm to improve the BEV perception method, and supplements the depth information of the two-dimensional image pixels on the basis of the two-dimensional image data with the depth information of the LiDAR point cloud. These pixels with supplemented depth information and the pixels with the depth information of the LiDAR point cloud are used together as training samples for the training of the BEV depth prediction model, so that the final BEV depth distribution prediction result has a higher degree of restoration.

[0046] Specifically, the process of using the square root 3 subdivision surface algorithm to supplement the depth information of the two-dimensional image pixel points is as follows:

[0047] First, the LiDAR point cloud is connected into a triangular mesh. The connection method is to connect any LiDAR point with the two closest LiDAR points to form a polyhedron composed of multiple triangles. Then, the surface of the polyhedron is subdivided using the root 3 subdivision surface algorithm. After multiple subdivisions, the mesh finally forms an approximate surface that is highly similar to the three-dimensional contours of all pixels in the two-dimensional image, so that the depth information of each pixel can be calculated. Then, the pixels are filtered according to certain rules, and the depth information of the filtered pixels is calculated, thereby obtaining a denser pixel depth distribution map with more depth information.

[0048] The calculation process of the square root 3 subdivision surface algorithm is as follows: each time a subdivision is performed, a new vertex is inserted into each triangle, and the new vertex is used as the center of gravity of the triangle. The new vertex is connected to the three vertices of the triangle to obtain three new triangles; finally, the three edges of the original triangle are removed to complete a subdivision. In this way, each subdivision triples the number of triangles before the subdivision, and two subdivisions make it nine times, and so on. After continuous subdivision, the added triangles become smaller and smaller, and the entire surface becomes more and more refined. Subdivision surfaces can continuously approach real surfaces, and the resolution and restoration of the surface can be improved by increasing the number of subdivisions.

[0049] Through the above-mentioned root 3 subdivision surface algorithm, an approximate surface mesh that is highly similar to the three-dimensional contours of all pixels of the two-dimensional image is obtained, which provides a basis for calculating and filtering the depth information of the pixel points. For each pixel point on the two-dimensional image, the triangle corresponding to it in the approximate surface mesh can be determined, and then the vertex information of the triangle and the position of the pixel point in the triangle are used to estimate the depth of the pixel point through interpolation methods (such as bilinear interpolation, bicubic interpolation, etc.). Directly calculating the depth information of all pixels may generate a large amount of redundant or inaccurate data. Therefore, these pixels need to be screened. There can be many screening rules, such as: based on the consistency of depth information (for example, if the depth of a pixel is very different from the depth of its neighboring pixels, it may be regarded as an outlier and excluded), or based on certain specific geometric or texture features. For the filtered pixels, their depth information can be further optimized or corrected to ensure that they are more accurate and continuous. The specific method for optimizing or correcting these depth information can be: Depth smoothing: Noise in the depth image may cause depth discontinuity. By applying a smoothing filter (such as a Gaussian filter), this discontinuity can be reduced and the accuracy of the depth information can be improved; Use edge information: The edges of the image are usually associated with sudden changes in depth. Image edge detection techniques (such as the Canny edge detector) can be used to guide depth correction so that depth changes are more consistent with object boundaries; Stereo matching or multi-view geometry: Use images from multiple perspectives to further refine depth information. By comparing pixels from different perspectives, a more accurate estimate of depth can be made. Finally, a depth distribution map is obtained, in which each pixel has depth information associated with it.

[0050] After obtaining the depth distribution of all pixels, it is converted into surround view BEV data of the scene, and the surround view BEV data is input into a trained CNN model for identifying and classifying the scene type to which the BEV scene data belongs, the identified scene type is obtained, the scene information is updated, and finally the scene understanding is completed.

[0051] Corresponding to the method for identifying scene types for autonomous driving described in the first embodiment of the present invention, the second embodiment of the present invention provides a device for identifying scene types for autonomous driving, including:

[0052] A map scene annotation recognition module is used to obtain map annotation data of a scene based on the vehicle positioning when the vehicle enters a scene, and obtain the scene type according to the map annotation data;

[0053] An image recognition module, for obtaining image data perceived for the scene and performing recognition to obtain a recognized scene type if map annotation data for the scene is not obtained;

[0054] The BEV recognition module is used to obtain the depth information of the two-dimensional image pixels through the pre-trained BEV depth prediction model if the confidence of image recognition does not meet the preset requirements, and form the depth distribution of the scene, and then convert the depth distribution into the BEV space to obtain the surround view BEV data of the scene; input the surround view BEV data into the pre-trained scene type recognition model to obtain the recognized scene type.

[0055] Preferably, when training the BEV depth prediction model, pixel points whose depth information is supplemented by a square root 3 subdivision surface algorithm and pixel points having depth information of a LiDAR point cloud are used as training samples.

[0056] Preferably, the method of supplementing the depth information by using a square root 3 subdivision surface algorithm specifically includes:

[0057] Connect any LiDAR point in the acquired LiDAR point cloud data to its two nearest LiDAR points to form a polyhedron composed of multiple triangular faces;

[0058] Subdividing the triangular surface of the polyhedron multiple times using a square root 3 subdivision surface algorithm to obtain an approximate surface mesh representing the three-dimensional contours of all pixel points of the two-dimensional image;

[0059] Depth information of each pixel is calculated based on the approximate surface mesh.

[0060] Preferably, the triangular surface of the polyhedron is subdivided multiple times using a square root 3 subdivision surface algorithm, specifically including: each time the subdivision is performed, a new vertex is inserted into each triangle, the new vertex is used as the center of gravity of the original triangle, the new vertex is connected to the three vertices of the original triangle, and three new triangles are obtained, and then the three sides of the original triangle are removed.

[0061] Preferably, calculating the depth information of each pixel point based on the approximate surface mesh specifically includes: determining the triangle corresponding to each pixel point in the approximate surface mesh, using the vertex information of the triangle and the position of the pixel point in the triangle, and estimating the depth information of the pixel point through an interpolation method.

[0062] Corresponding to the scene type identification method for autonomous driving described in the aforementioned embodiment 1 of the present invention, embodiment 3 of the present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored computer program; wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the scene type identification method for autonomous driving as described in embodiment 1 of the present invention.

[0063] Preferably, the computer program can be divided into one or more modules / units (such as computer program 1, computer program 2, ...), and the one or more modules / units are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments that can complete specific functions, and the instruction segments are used to describe the execution process of the computer program in the device.

[0064] The processor may be a central processing unit (CPU), or other general-purpose processors, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor, or the processor may be any conventional processor. The processor is the control center of the device, and various parts of the device are connected using various interfaces and lines.

[0065] The memory mainly includes a program storage area and a data storage area, wherein the program storage area can store an operating system, an application required for at least one function, etc., and the data storage area can store related data, etc. In addition, the memory can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard disk, a smart memory card (SmartMedia Card, SMC), a secure digital (Secure Digital, SD) card, and a flash card (Flash Card), etc., or the memory can also be other volatile solid-state storage devices.

[0066] It should be noted that the above-mentioned device may include but is not limited to a processor and a memory, which can be understood by those skilled in the art.

[0067] For the working principle and process of the above embodiment, please refer to the description of the above embodiment of the present invention, which will not be repeated here.

[0068] It can be seen from the above description that compared with the prior art, the beneficial effects of the present invention are: using the three mechanisms of map scene annotation, image recognition, and BEV perception to realize scene type recognition in two cases where there is map scene annotation information and no map or light map scene annotation information, thereby improving the reliability of scene recognition; at the same time, the scene understanding function is realized with the minimum computing power as much as possible through the three mechanisms in series, ensuring that the scene understanding function can generate scene attribute results to the greatest extent. The BEV perception method is improved, and the pixel depth information of the two-dimensional image is supplemented by the square root 3 subdivision surface algorithm, so that the final BEV depth distribution prediction result has a higher degree of restoration. The present invention can provide scene category information for the decision-making planning module, so that the decision-making planning module can better complete tasks in specific scenarios, improve point-to-point autonomous driving capabilities, and provide the prerequisites and cognitive basis for the "three-domain integration" of highway, city, and parking scenarios, thereby improving the user's riding experience.

[0069] The above disclosure is only the preferred embodiment of the present invention, which certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for identifying scene types for autonomous driving, characterized in that: include: When the vehicle enters a certain scene, the map annotation data of the scene is obtained based on the vehicle positioning, and the scene type is obtained according to the map annotation data; If the map annotation data of the scene is not obtained, image data perceived by the scene is obtained and recognized to obtain the recognized scene type; If the confidence of image recognition does not meet the preset requirements, the image data is passed through a pre-trained BEV depth prediction model to obtain the depth information of the two-dimensional image pixels, and the depth distribution of the scene is formed, and then the depth distribution is converted into the BEV space to obtain the surround view BEV data of the scene; the surround view BEV data is input into a pre-trained scene type recognition model to obtain the recognized scene type.

2. The method according to claim 1, characterized in that When training the BEV depth prediction model, pixel points whose depth information is supplemented by the square root 3 subdivision surface algorithm and pixel points with the depth information of the LiDAR point cloud are used as training samples.

3. The method according to claim 2, characterized in that The method of supplementing the depth information by using the square root 3 subdivision surface algorithm specifically includes: Connect any LiDAR point in the acquired LiDAR point cloud data to its two nearest LiDAR points to form a polyhedron composed of multiple triangular faces; Subdividing the triangular surface of the polyhedron multiple times using a square root 3 subdivision surface algorithm to obtain an approximate surface mesh representing the three-dimensional contours of all pixel points of the two-dimensional image; Depth information of each pixel is calculated based on the approximate surface mesh.

4. The method according to claim 3, characterized in that: The triangular surface of the polyhedron is subdivided multiple times using a square root 3 subdivision surface algorithm, specifically including: each time the subdivision is performed, a new vertex is inserted into each triangle, the new vertex is used as the center of gravity of the original triangle, the new vertex is connected to the three vertices of the original triangle, and three new triangles are obtained, and then the three sides of the original triangle are removed.

5. The method according to claim 3, characterized in that: The method of calculating the depth information of each pixel point based on the approximate surface mesh specifically includes: determining the triangle corresponding to each pixel point in the approximate surface mesh, using the vertex information of the triangle and the position of the pixel point in the triangle, and estimating the depth information of the pixel point through an interpolation method.

6. A scene type recognition device for autonomous driving, characterized in that: include: A map scene annotation recognition module is used to obtain map annotation data of a scene based on the vehicle positioning when the vehicle enters a scene, and obtain the scene type according to the map annotation data; An image recognition module, for obtaining image data perceived for the scene and performing recognition to obtain a recognized scene type if map annotation data for the scene is not obtained; The BEV recognition module is used to obtain the depth information of the two-dimensional image pixels through the pre-trained BEV depth prediction model if the confidence of image recognition does not meet the preset requirements, and form the depth distribution of the scene, and then convert the depth distribution into the BEV space to obtain the surround view BEV data of the scene; input the surround view BEV data into the pre-trained scene type recognition model to obtain the recognized scene type.

7. The device according to claim 6, characterized in that When training the BEV depth prediction model, pixel points whose depth information is supplemented by the square root 3 subdivision surface algorithm and pixel points with the depth information of the LiDAR point cloud are used as training samples.

8. The device according to claim 7, characterized in that The method of supplementing the depth information by using the square root 3 subdivision surface algorithm specifically includes: Connect any LiDAR point in the acquired LiDAR point cloud data to its two nearest LiDAR points to form a polyhedron composed of multiple triangular faces; Subdividing the triangular surface of the polyhedron multiple times using a square root 3 subdivision surface algorithm to obtain an approximate surface mesh representing the three-dimensional contours of all pixel points of the two-dimensional image; Depth information of each pixel is calculated based on the approximate surface mesh.

9. The device according to claim 8, characterized in that The triangular surface of the polyhedron is subdivided multiple times using a square root 3 subdivision surface algorithm, specifically including: each time the subdivision is performed, a new vertex is inserted into each triangle, the new vertex is used as the center of gravity of the original triangle, the new vertex is connected to the three vertices of the original triangle, and three new triangles are obtained, and then the three sides of the original triangle are removed.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program; wherein, when the computer program is running, it controls the device where the computer-readable storage medium is located to execute the scene type recognition method for autonomous driving as described in any one of claims 1 to 5.

Images