Automatic parking path planning method and device based on 3D GS

Abstract

The embodiment of the invention discloses an automatic parking path planning method and device based on 3D GS, and the method comprises the steps: sensing the information of the environment around a vehicle through a sensing sensor, and carrying out the preprocessing of the information; performing three-dimensional reconstruction by using a preset 3D GS algorithm and the preprocessed environment information to obtain a three-dimensional scene graph, dividing the three-dimensional scene graph into voxel grids with a uniform size according to a predetermined size of a grid map, and enabling all the voxel grids to correspond to pixel positions of the grid map one by one according to occupied and unoccupied voxel grids, so as to obtain a three-dimensional scene graph; generating a final grid map; and performing path planning in the final grid map according to the starting point pose and the terminal point pose of the vehicle to obtain a parking path. According to the method, the 3D GS algorithm is combined with a traditional perception sensor, in the face of a complex environment, environment information can be better obtained, a more accurate scene map can be obtained, and therefore a parking path can be planned more flexibly and more accurately.

Description

Technical Field

[0001] The present application relates to the technical field of automatic parking, and in particular to a method and device for automatic parking path planning based on 3D GS. Background Art

[0002] The goal of automatic parking is to obtain information about the surrounding environment through sensors such as cameras and radars, plan the driving path through algorithms, and control the vehicle to complete the parking action. Existing parking devices mainly rely on traditional perception sensors such as cameras and radars to collect scene information, and then use path planning algorithms to plan and find the best path for automatic parking.

[0003] However, the perception sensors used in traditional parking are greatly affected by environmental factors such as weather and magnetic fields, and their ability to understand scenes is limited, and they may even fail to work properly in certain environments. In addition, due to the limitations of the perception capabilities of traditional parking, the efficiency of its path planning algorithm is low, and it is difficult to quickly find the best path, especially in complex environments, and it takes a long time to calculate, thus affecting parking efficiency. Summary of the invention

[0004] This application provides an automatic parking path planning method and device based on 3D GS to improve the efficiency and accuracy of automatic parking. The specific scheme is as follows:

[0005] In a first aspect, an automatic parking path planning method based on 3D GS is characterized by comprising:

[0006] The environmental information around the vehicle is sensed by the perception sensor, and preprocessed to obtain preprocessed environmental information;

[0007] Based on the preprocessed environmental information, a 3D scene graph is obtained by 3D reconstruction using a preset 3D Gaussian Splatting algorithm;

[0008] Dividing the three-dimensional scene graph according to a predetermined size of the grid map to obtain voxel grids of uniform size, and traversing all voxel grids to determine occupied voxel grids and unoccupied voxel grids;

[0009] According to the occupied voxel grids and the unoccupied voxel grids, all voxel grids are mapped one by one to the pixel positions of the grid map to generate the final grid map;

[0010] In the final grid map, the path planning is performed according to the starting position and the ending position of the vehicle to obtain the parking path, and the vehicle is controlled according to the parking path to complete automatic parking.

[0011] Optionally, the environmental information around the vehicle is sensed by a perception sensor, and preprocessed to obtain the preprocessed environmental information including:

[0012] The environmental information around the vehicle is sensed by the perception sensor, and the environmental information includes the scene image around the vehicle and the point cloud data of obstacles;

[0013] The scene image and the point cloud data are preprocessed according to the corresponding preprocessing method to obtain the preprocessed scene image and the preprocessed point cloud data.

[0014] Optionally, the scene image and the point cloud data are preprocessed according to a corresponding preprocessing method, and the preprocessed scene image and the preprocessed point cloud data are obtained, including:

[0015] Perform image enhancement and feature extraction enhancement on the scene image to obtain a preprocessed scene image;

[0016] The point cloud data is denoised and registered to obtain the preprocessed point cloud data.

[0017] Optionally, based on the preprocessed environment information, a 3D scene graph is obtained by performing 3D reconstruction using a preset 3D Gaussian Splatting algorithm, including:

[0018] The preprocessed environmental information is input into the preset 3D Gaussian Splatting algorithm, and the preprocessed scene image is subjected to colmap, observation transformation, projection transformation, viewport transformation, rasterization and rendering in sequence. After rendering, a rendered image is obtained, and a loss value between the rendered image and the input image is returned. The loss value is used to adaptively change the point cloud distribution mode to obtain a three-dimensional scene graph.

[0019] Optionally, the predetermined size of the grid map includes the size, resolution and voxel size of the grid map.

[0020] Optionally, dividing the three-dimensional scene graph according to a predetermined size of the grid map to obtain voxel grids of uniform size, and traversing all voxel grids to determine occupied voxel grids and unoccupied voxel grids includes:

[0021] Divide each dimension of the three-dimensional scene graph into a voxel grid of uniform size according to the size, resolution and voxel size of the grid map;

[0022] Among them, the size of each voxel grid is the same as the voxel size, the resolution of each voxel grid is the same as the resolution size, and the total size of all voxel grids in each dimension is the same as the size of the grid map;

[0023] For each voxel grid, it is determined whether the voxel grid contains point cloud data of obstacles. If so, the voxel grid is considered to be occupied; otherwise, the voxel grid is considered to be unoccupied.

[0024] Optionally, according to the occupied voxel grids and the unoccupied voxel grids, all voxel grids are mapped one by one to the pixel positions of the grid map to generate the final grid map, including:

[0025] The occupied voxel grids are considered as obstacles on the grid map, and the unoccupied voxel grids are considered as traversable areas;

[0026] According to the occupied voxel grids and the unoccupied voxel grids, all voxel grids are mapped one by one to the pixel positions on the grid map to generate the final grid map.

[0027] Optionally, performing path planning according to the starting position and the ending position of the vehicle in the final grid map to obtain a parking path, and controlling the vehicle to complete automatic parking according to the parking path includes:

[0028] In the final grid map, determine the starting position and ending position of the vehicle;

[0029] By integrating the TEB algorithm with the Hybrid A* algorithm, and based on the starting and ending positions, the vehicle path is planned to obtain the parking path of the vehicle;

[0030] The parking path is input into the vehicle controller to control the vehicle to complete automatic parking according to the parking path.

[0031] Optionally, after the vehicle is controlled to complete automatic parking according to the parking path, the automatic parking path planning method based on 3D GS further includes:

[0032] The starting position, end position and parking path are input into the model predictive controller, and the vehicle is tracked in real time to confirm whether the vehicle completes automatic parking according to the parking path.

[0033] In a second aspect, the present application provides an automatic parking path planning device based on 3D GS, including an environmental information perception module, a three-dimensional map reconstruction module, a voxel grid division module, a grid map generation module and a parking path planning module;

[0034] The environmental information perception module is configured to perceive the environmental information around the vehicle through a perception sensor and preprocess the environmental information to obtain preprocessed environmental information;

[0035] The three-dimensional map reconstruction module is configured to perform three-dimensional reconstruction based on the pre-processed environmental information and use a preset 3D Gaussian Splatting algorithm to obtain a three-dimensional scene graph;

[0036] A voxel grid partitioning module is configured to partition the three-dimensional scene graph according to a predetermined size of the grid map to obtain voxel grids of uniform size, and traverse all voxel grids to determine occupied voxel grids and unoccupied voxel grids;

[0037] A grid map generation module is configured to correspond all voxel grids to pixel positions of the grid map one by one according to occupied voxel grids and unoccupied voxel grids to generate a final grid map;

[0038] The parking path planning module is configured to perform path planning according to the starting position and the ending position of the vehicle in the final grid map to obtain a parking path, and control the vehicle to complete automatic parking according to the parking path.

[0039] The innovative features of the embodiments of the present application include:

[0040] The embodiment of the present application discloses a method and device for automatic parking path planning based on 3D GS, which senses environmental information around a vehicle through a sensing sensor and preprocesses the environmental information to obtain preprocessed environmental information; based on the preprocessed environmental information, a three-dimensional reconstruction is performed using a preset 3D Gaussian Splatting algorithm to obtain a three-dimensional scene graph; the three-dimensional scene graph is divided according to a predetermined size of a grid map to obtain voxel grids of a uniform size, and all voxel grids are traversed to determine occupied voxel grids and unoccupied voxel grids; according to the occupied voxel grids and unoccupied voxel grids, all voxel grids are mapped one by one to pixel positions of the grid map to generate a final grid map; in the final grid map, a path is planned according to the starting position and the end position of the vehicle to obtain a parking path, and the vehicle is controlled according to the parking path to complete automatic parking. This application combines the 3D Gaussian Splatting algorithm with traditional perception sensors. When facing complex environments, it can better obtain environmental information and obtain a more accurate scene map, so that the subsequent planning system can perform more flexible and accurate path planning; in addition, this application can greatly enhance the anti-interference ability of the vehicle perception system, making the perception system more redundant and robust, thereby greatly improving parking efficiency and parking safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art description are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative work.

[0042] Figure 1 A schematic diagram of a flow chart of an automatic parking path planning method based on 3D GS provided in an embodiment of the present application;

[0043] Figure 2 A detailed schematic diagram of an automatic parking path planning method based on 3D GS provided in an embodiment of the present application;

[0044] Figure 3 A schematic diagram of the structure of an automatic parking path planning device based on 3D GS provided in an embodiment of the present application. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0046] It should be noted that the terms "including" and "having" and any variations thereof in the embodiments of the present application and the accompanying drawings are intended to cover non-exclusive inclusions. For example, a process, method, device, product or equipment comprising a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or equipment.

[0047] Due to the limitations of the perception capabilities of traditional parking, its path planning algorithm is inefficient and it is difficult to quickly find the best path, especially in complex environments. It takes a long time to calculate, which affects parking efficiency. This application uses a more sophisticated and accurate three-dimensional scene model to quickly identify obstacles and feasible areas, achieve more sophisticated and faster path planning, and improve parking efficiency.

[0048] The embodiment of the present application discloses a method and device for automatic parking path planning based on 3D GS, which senses environmental information around a vehicle through a sensing sensor and preprocesses the environmental information to obtain preprocessed environmental information; based on the preprocessed environmental information, a three-dimensional reconstruction is performed using a preset 3D Gaussian Splatting algorithm to obtain a three-dimensional scene graph; the three-dimensional scene graph is divided according to a predetermined size of a grid map to obtain voxel grids of a uniform size, and all voxel grids are traversed to determine occupied voxel grids and unoccupied voxel grids; according to the occupied voxel grids and unoccupied voxel grids, all voxel grids are mapped one by one to pixel positions of the grid map to generate a final grid map; in the final grid map, a path is planned according to the starting position and the end position of the vehicle to obtain a parking path, and the vehicle is controlled according to the parking path to complete automatic parking. This application combines the 3D Gaussian Splatting algorithm with traditional perception sensors. When facing complex environments, it can better obtain environmental information and obtain a more accurate scene map, so that the subsequent planning system can perform more flexible and accurate path planning; in addition, this application can greatly enhance the anti-interference ability of the vehicle perception system, making the perception system more redundant and robust, thereby greatly improving parking efficiency and parking safety.

[0049] The embodiments of the present application are described in detail below.

[0050] Figure 1 A schematic diagram of a process flow of an automatic parking path planning method based on 3D GS provided in an embodiment of the present application is shown. Figure 2 Detailed schematic diagram of an automatic parking path planning method based on 3D GS provided in an embodiment of the present application. The purpose of the present application is to plan an automatic parking path.

[0051] Combination Figure 1 and Figure 2 , the present application provides an automatic parking path planning method based on 3D GS, comprising:

[0052] S1, sensing the environmental information around the vehicle through the sensing sensor, and preprocessing it to obtain preprocessed environmental information;

[0053] In this step, various types of perception sensors such as cameras or radars can be used to perceive the environmental information around the vehicle, which includes images of the surrounding scene and point cloud data of obstacles.

[0054] S2, based on the preprocessed environmental information, a preset 3D Gaussian Splatting algorithm is used to perform 3D reconstruction to obtain a 3D scene graph;

[0055] In this step, the pre-processed environment information can be input into the preset 3D Gaussian Splatting (3DGS) algorithm, and the pre-processed scene image is subjected to colmap, observation transformation, projection transformation, viewport transformation, rasterization and rendering in sequence. After the rendering is completed, a rendered image is obtained, and the loss value between the rendered image and the input image (original image) is returned. The loss value is used to adaptively change the point cloud distribution mode to obtain a three-dimensional scene graph. The input image is the image contained in the pre-processed environment information.

[0056] The present application may input the preprocessed scene image alone into the 3D Gaussian Splatting algorithm, or may input the preprocessed scene image and point cloud data into the 3D Gaussian Splatting algorithm together.

[0057] It is worth noting that the input of the 3D Gaussian Splatting algorithm is a picture or a video, and then the point cloud is obtained through colmap. If the point cloud data is directly input, it is equivalent to skipping the colmap step and directly performing operations such as observation transformation and projection transformation. The first solution can input the pre-processed scene picture and sequentially undergo colmap, observation transformation, projection transformation, viewport transformation rasterization and rendering. After the rendering is completed, the rendered image is obtained, and the loss value of the rendered image and the input image (original image) is returned. The loss value is used to adaptively change the point cloud distribution mode to obtain a three-dimensional scene graph. The second solution adds point cloud data to take into account two points: 1. If you only rely on the camera to perceive environmental information, it may be interfered by some external forces. Adding radar to obtain point cloud data can prevent this situation and make the data more reliable; 2. The camera captures environmental information, and the radar also extracts the point cloud data of obstacles and the point cloud data of the environment, which can be compared with each other to enhance the perception of the environment. Afterwards, the present application can input the preprocessed scene image and point cloud data into the 3D Gaussian Splatting algorithm together, which can improve the accuracy when reconstructing the three-dimensional scene graph.

[0058] The 3D Gaussian Splatting algorithm performs observation transformation, projection transformation, viewport transformation, rasterization and rendering in sequence. The observation transformation and projection transformation convert the three-dimensional Gaussian ellipsoid to a two-dimensional plane, so the object of observation transformation and projection transformation is a three-dimensional Gaussian ellipsoid. The viewport transformation stretches the 2D Gaussian to the viewport size and position suitable for rendering, so as to ensure that the 2D Gaussian can be correctly aligned in the final rendered image. Rasterization converts the graphic object into pixels to facilitate the subsequent rendering of pixels, that is, the object of viewport transformation and rasterization is the 2D Gaussian. The object of rendering is the pixel obtained by rasterization. The input object of the 3D Gaussian Splatting algorithm is a picture or video, so as to obtain a point cloud, and the point cloud generates a three-dimensional Gaussian ellipsoid. First, the three-dimensional Gaussian ellipsoid is projected onto a two-dimensional plane, then rendered and compared with the original two-dimensional picture for loss, and finally the distribution of the point cloud is changed by the gradient backpropagation method to obtain a three-dimensional scene graph.

[0059] S3, dividing the three-dimensional scene graph according to a predetermined size of the grid map to obtain voxel grids of uniform size, and traversing all voxel grids to determine occupied voxel grids and unoccupied voxel grids;

[0060] The predetermined size of the grid map includes the size, resolution and voxel size of the grid map.

[0061] It is worth noting that this step requires dividing the three-dimensional scene graph according to the predetermined size of the grid map, and dividing the three-dimensional scene graph into multiple voxel grids. Each voxel grid, resolution and voxel size are the same as the voxel size and resolution of the predetermined grid map. All voxel grids in each dimension constitute the size of the grid map.

[0062] S4, according to the occupied voxel grids and the unoccupied voxel grids, all voxel grids are mapped one by one to the pixel positions of the grid map to generate a final grid map;

[0063] It is worth noting that this step requires that all voxel grids be mapped one by one to the pixel positions on the grid map, and the two types of voxel grids obtained in step 3 are distinguished. The occupied pixels on the corresponding grid map are regarded as obstacles, and the unoccupied pixels on the corresponding grid map are regarded as passable areas. After completing the correspondence of all voxel grids, the final grid map is generated.

[0064] S5, performing path planning according to the starting position and the ending position of the vehicle in the final grid map to obtain a parking path, and controlling the vehicle to complete automatic parking according to the parking path.

[0065] It is worth noting that in this step, the final grid map can be input into the computing device, and then the starting position and the end position of the vehicle are set in the final grid map, and the parking path is planned using the path planning algorithm. The parking path is then input into the vehicle controller so that the vehicle automatically completes parking according to the parking path.

[0066] refer to Figure 2 ,from Figure 2 The sensing sensor senses the environmental information, obtains the scene image and point cloud data, and then performs data preprocessing to obtain the preprocessed data, which is input into the 3D Gaussian Splatting algorithm to output the three-dimensional scene graph. After determining the size, resolution and voxel size of the grid map, the three-dimensional scene graph is divided into voxel grids, and all voxel grids are traversed to determine the voxel grids containing the point cloud data of the obstacle, and the voxel grid is marked as occupied; if the point cloud data does not contain the obstacle, the voxel grid is determined to be unoccupied. Then, according to the pixels corresponding to the grid map that are occupied or unoccupied, the unoccupied voxel grids are marked as passable areas on the grid map, and the occupied voxel grids are marked as inaccessible areas, and the final grid map is obtained. Then, the grid map and the model predictive controller are used to plan the parking path, and finally the parking operation is completed for the vehicle to be parked.

[0067] In an optional implementation of the present application, sensing the environmental information around the vehicle by a sensing sensor, and preprocessing the environmental information to obtain the preprocessed environmental information includes:

[0068] The environmental information around the vehicle is sensed by the perception sensor, and the environmental information includes the scene image around the vehicle and the point cloud data of obstacles;

[0069] The scene image and the point cloud data are preprocessed according to the corresponding preprocessing method to obtain the preprocessed scene image and the preprocessed point cloud data.

[0070] This implementation requires preprocessing of scene images and point cloud data in order to improve data quality and facilitate subsequent identification and capture of targets.

[0071] In an optional implementation manner of the present application, the scene image and the point cloud data are preprocessed according to a corresponding preprocessing method, and the preprocessed scene image and the preprocessed point cloud data are obtained, including:

[0072] Perform image enhancement and feature extraction enhancement on the scene image to obtain a preprocessed scene image;

[0073] The point cloud data is denoised and registered to obtain the preprocessed point cloud data.

[0074] It is worth noting that this implementation requires preprocessing of the collected environmental information. For camera data, i.e., images, it can be processed by methods such as image enhancement and feature extraction. Feature extraction refers to the extraction of color features, texture features, property features, etc., to facilitate the subsequent recognition and capture of image information. Radar data, i.e., point cloud data, can be processed by methods such as point cloud denoising and registration to improve data quality and provide more accurate input data for the subsequent 3D Gaussian Splatting algorithm.

[0075] In an optional implementation of the present application, dividing the three-dimensional scene graph according to a predetermined size of the grid map to obtain voxel grids of uniform size, and traversing all voxel grids to determine occupied voxel grids and unoccupied voxel grids includes:

[0076] Divide each dimension of the three-dimensional scene graph into a voxel grid of uniform size according to the size, resolution and voxel size of the grid map;

[0077] Among them, the size of each voxel grid is the same as the voxel size, the resolution of each voxel grid is the same as the resolution size, and the total size of all voxel grids in each dimension is the same as the size of the grid map;

[0078] For each voxel grid, it is determined whether the voxel grid contains point cloud data of obstacles. If so, the voxel grid is considered to be occupied; otherwise, the voxel grid is considered to be unoccupied.

[0079] It is worth noting that whether each voxel grid contains point cloud data of obstacles can determine whether the voxel grid is occupied. Similarly, if the adjacent voxel grid is occupied, it will be considered as an inaccessible area on the grid map, which can improve the generation efficiency when forming the grid map.

[0080] In an optional implementation of the present application, according to the occupied voxel grids and the unoccupied voxel grids, all voxel grids are mapped one by one to the pixel positions of the grid map to generate the final grid map, including:

[0081] The occupied voxel grids are considered as obstacles on the grid map, and the unoccupied voxel grids are considered as traversable areas;

[0082] According to the occupied voxel grids and the unoccupied voxel grids, all voxel grids are mapped one by one to the pixel positions on the grid map to generate the final grid map.

[0083] Exemplarily, after determining the size of the grid map, that is, the size, resolution size and voxel size, the entire three-dimensional scene graph is divided. For example, if the grid map is 100*100, the voxel grid of the three-dimensional scene graph is divided into 1*1*1. In this way, 100 voxel grids in the x, y, and z directions of the three-dimensional scene can be obtained. Then, it is determined whether there is obstacle point cloud data in the voxel grid, which corresponds to whether this position in the grid map is a passable area.

[0084] In an optional implementation of the present application, path planning is performed according to the starting position and the ending position of the vehicle in the final grid map to obtain a parking path, and the vehicle is controlled according to the parking path to complete automatic parking, including:

[0085] In the final grid map, determine the starting position and ending position of the vehicle;

[0086] In this step, the final grid map can be input into the computing device and displayed. The starting position and the ending position of the vehicle are obtained by the vehicle controller, and then displayed on the final grid map.

[0087] By integrating the TEB algorithm with the Hybrid A* algorithm, and based on the starting and ending positions, the vehicle path is planned to obtain the parking path of the vehicle;

[0088] The parking path is input into the vehicle controller to control the vehicle to complete automatic parking according to the parking path.

[0089] In an optional implementation of the present application, after the vehicle is controlled to complete automatic parking according to the parking path, the automatic parking path planning method based on 3D GS further includes:

[0090] The starting position, end position and parking path are input into the model predictive controller, and the vehicle is tracked in real time to confirm whether the vehicle completes automatic parking according to the parking path.

[0091] like Figure 3 As shown, the present application discloses an automatic parking path planning device based on 3D GS, including an environmental information perception module 31, a three-dimensional map reconstruction module 32, a voxel grid division module 33, a grid map generation module 34 and a parking path planning module 35;

[0092] The environmental information sensing module 31 is configured to sense the environmental information around the vehicle through a sensing sensor and preprocess the environmental information to obtain preprocessed environmental information;

[0093] The three-dimensional map reconstruction module 32 is configured to perform three-dimensional reconstruction based on the pre-processed environmental information and use a preset 3D Gaussian Splatting algorithm to obtain a three-dimensional scene map;

[0094] The voxel grid division module 33 is configured to divide the three-dimensional scene graph according to a predetermined size of the grid map to obtain voxel grids of uniform size, and traverse all voxel grids to determine occupied voxel grids and unoccupied voxel grids;

[0095] The grid map generation module 34 is configured to correspond all voxel grids to pixel positions of the grid map one by one according to the occupied voxel grids and the unoccupied voxel grids, so as to generate a final grid map;

[0096] The parking path planning module 35 is configured to perform path planning according to the starting position and the ending position of the vehicle in the final grid map to obtain a parking path, and control the vehicle to complete automatic parking according to the parking path.

[0097] Optionally, the environmental information perception module 31 is configured as follows:

[0098] The environmental information around the vehicle is sensed by the perception sensor, and the environmental information includes the scene image around the vehicle and the point cloud data of obstacles;

[0099] The scene image and the point cloud data are preprocessed according to the corresponding preprocessing method to obtain the preprocessed scene image and the preprocessed point cloud data.

[0100] Optionally, the environmental information perception module 31 is configured as follows:

[0101] Perform image enhancement and feature extraction enhancement on the scene image to obtain a preprocessed scene image;

[0102] The point cloud data is denoised and registered to obtain the preprocessed point cloud data.

[0103] Optionally, the three-dimensional map reconstruction module 32 is configured to:

[0104] The preprocessed environmental information is input into the preset 3D Gaussian Splatting algorithm, and the preprocessed scene image is subjected to colmap, observation transformation, projection transformation, viewport transformation, rasterization and rendering in sequence. After rendering, a rendered image is obtained, and a loss value between the rendered image and the input image is returned. The loss value is used to adaptively change the point cloud distribution mode to obtain a three-dimensional scene graph.

[0105] The predetermined size of the grid map includes the size, resolution and voxel size of the grid map.

[0106] Optionally, the voxel grid division module 33 is configured to:

[0107] Divide each dimension of the three-dimensional scene graph into a voxel grid of uniform size according to the size, resolution and voxel size of the grid map;

[0108] Among them, the size of each voxel grid is the same as the voxel size, the resolution of each voxel grid is the same as the resolution size, and the total size of all voxel grids in each dimension is the same as the size of the grid map;

[0109] For each voxel grid, it is determined whether the voxel grid contains point cloud data of obstacles. If so, the voxel grid is considered to be occupied; otherwise, the voxel grid is considered to be unoccupied.

[0110] Optionally, the grid map generation module 34 is configured to:

[0111] The occupied voxel grids are considered as obstacles on the grid map, and the unoccupied voxel grids are considered as traversable areas;

[0112] According to the occupied voxel grids and the unoccupied voxel grids, all voxel grids are mapped one by one to the pixel positions on the grid map to generate the final grid map.

[0113] Optionally, the parking path planning module 35 is configured to:

[0114] In the final grid map, determine the starting position and ending position of the vehicle;

[0115] By integrating the TEB algorithm with the Hybrid A* algorithm, and based on the starting and ending positions, the vehicle path is planned to obtain the parking path of the vehicle;

[0116] The parking path is input into the vehicle controller to control the vehicle to complete automatic parking according to the parking path.

[0117] Optionally, the present application discloses an automatic parking path planning device based on 3D GS, further comprising a real-time tracking module, configured as follows:

[0118] The starting position, end position and parking path are input into the model predictive controller, and the vehicle is tracked in real time to confirm whether the vehicle completes automatic parking according to the parking path.

[0119] Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of one embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present application.

[0120] Those skilled in the art can understand that the modules in the device in the embodiment can be distributed in the device in the embodiment according to the description of the embodiment, or can be changed accordingly and located in one or more devices different from the embodiment. The modules in the above embodiment can be combined into one module, or can be further divided into multiple sub-modules.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An automatic parking path planning method based on 3D GS, characterized in that: include: The environmental information around the vehicle is sensed by the sensing sensor, and preprocessed to obtain preprocessed environmental information; Based on the preprocessed environmental information, a preset 3D Gaussian Splatting algorithm is used to perform three-dimensional reconstruction to obtain a three-dimensional scene graph; Dividing the three-dimensional scene graph according to a predetermined size of the grid map to obtain voxel grids of uniform size, and traversing all voxel grids to determine occupied voxel grids and unoccupied voxel grids; According to the occupied voxel grids and the unoccupied voxel grids, all voxel grids are mapped one by one to the pixel positions of the grid map to generate the final grid map; In the final grid map, a parking path is obtained by performing path planning according to the starting position and the ending position of the vehicle, and the vehicle is controlled to complete automatic parking according to the parking path.

2. The automatic parking path planning method based on 3D GS according to claim 1, characterized in that: The sensing of the environmental information around the vehicle by the sensing sensor and preprocessing the environmental information to obtain the preprocessed environmental information include: Perceiving environmental information around the vehicle through a perception sensor, wherein the environmental information includes a scene image around the vehicle and point cloud data of obstacles; The scene image and the point cloud data are preprocessed according to corresponding preprocessing methods to obtain a preprocessed scene image and preprocessed point cloud data.

3. The automatic parking path planning method based on 3D GS according to claim 2 is characterized in that: The preprocessing of the scene image and the point cloud data according to the corresponding preprocessing method to obtain the preprocessed scene image and the preprocessed point cloud data includes: Performing image enhancement and feature extraction enhancement on the scene image to obtain a preprocessed scene image; The point cloud data is denoised and point cloud registered to obtain preprocessed point cloud data.

4. The automatic parking path planning method based on 3D GS according to claim 1, characterized in that: The three-dimensional reconstruction based on the pre-processed environmental information and using a preset 3D Gaussian Splatting algorithm to obtain a three-dimensional scene graph includes: The preprocessed environmental information is input into the preset 3D Gaussian Splatting algorithm, and the preprocessed scene image is subjected to colmap, observation transformation, projection transformation, viewport transformation, rasterization and rendering in sequence. After rendering, a rendered image is obtained, and a loss value between the rendered image and the input image is returned. The loss value is used to adaptively change the point cloud distribution mode to obtain a three-dimensional scene graph.

5. The automatic parking path planning method based on 3D GS according to claim 1, characterized in that: The predetermined size of the grid map includes the size, resolution and voxel size of the grid map.

6. The automatic parking path planning method based on 3D GS according to claim 5, characterized in that: The step of dividing the three-dimensional scene graph according to a predetermined size of the grid map to obtain voxel grids of uniform size, and traversing all voxel grids to determine occupied voxel grids and unoccupied voxel grids includes: Dividing each dimension of the three-dimensional scene graph into a voxel grid of uniform size according to the size, resolution size and voxel size of the grid map; wherein the size of each voxel grid is the same as the voxel size, the resolution of each voxel grid is the same as the resolution size, and the total size of all voxel grids in each dimension is the same as the size of the grid map; For each voxel grid, it is determined whether the voxel grid contains point cloud data of obstacles. If so, the voxel grid is considered to be occupied; otherwise, the voxel grid is considered to be unoccupied.

7. The automatic parking path planning method based on 3D GS according to claim 1, characterized in that: The step of corresponding all voxel grids to pixel positions of the grid map one by one according to the occupied voxel grids and the unoccupied voxel grids to generate the final grid map includes: The occupied voxel grids are considered as obstacles on the grid map, and the unoccupied voxel grids are considered as traversable areas; According to the occupied voxel grids and the unoccupied voxel grids, all voxel grids are mapped one by one to the pixel positions on the grid map to generate the final grid map.

8. The automatic parking path planning method based on 3D GS according to claim 1, characterized in that: The step of performing path planning according to the starting position and the ending position of the vehicle in the final grid map to obtain a parking path, and controlling the vehicle to complete automatic parking according to the parking path includes: In the final grid map, determining the starting position and the ending position of the vehicle; By fusing the TEB algorithm with the Hybrid A* algorithm, and according to the starting point posture and the end point posture, the vehicle path is planned to obtain the parking path of the vehicle; The parking path is input into a vehicle controller to control the vehicle to complete automatic parking according to the parking path.

9. The automatic parking path planning method based on 3D GS according to claim 1, characterized in that: After the vehicle is controlled to complete automatic parking according to the parking path, the automatic parking path planning method based on 3DGS further includes: The starting point posture, the end point posture and the parking path are input into a model prediction controller, and the vehicle is tracked in real time to confirm whether the vehicle completes automatic parking according to the parking path.

10. An automatic parking path planning device based on 3D GS, characterized in that: It includes environment information perception module, 3D map reconstruction module, voxel grid division module, grid map generation module and parking path planning module; The environmental information perception module is configured to perceive the environmental information around the vehicle through a perception sensor and preprocess the environmental information to obtain preprocessed environmental information; A three-dimensional map reconstruction module is configured to perform three-dimensional reconstruction based on the pre-processed environmental information and use a preset 3D Gaussian Splatting algorithm to obtain a three-dimensional scene graph; A voxel grid partitioning module is configured to partition the three-dimensional scene graph according to a predetermined size of the grid map to obtain voxel grids of uniform size, and traverse all voxel grids to determine occupied voxel grids and unoccupied voxel grids; A grid map generation module is configured to correspond all voxel grids to pixel positions of the grid map one by one according to occupied voxel grids and unoccupied voxel grids to generate a final grid map; The parking path planning module is configured to perform path planning in the final grid map according to the starting position and the ending position of the vehicle to obtain a parking path, and control the vehicle to complete automatic parking according to the parking path.

Images