Automatic parking obstacle avoidance method based on vehicle-mounted image data

Abstract

The present invention discloses an automatic parking obstacle avoidance method based on vehicle-mounted image data, which relates to the field of automatic parking technology and effectively improves the safety and reliability of automatic parking. The present invention establishes a dynamic parking scene model centered on the vehicle through multi-perspective scene images, divides a number of scene object models in the dynamic parking scene model, and sets scene semantic labels for each scene object model. The spatial position of the scene object model in the next scene model update cycle is obtained. Based on the spatial position of the scene object model in the next scene model update cycle, it is determined whether a collision-free parking path will cause a collision. Based on the judgment result, multiple detour collision-free parking paths are generated, and the real-time risk value of each detour collision-free parking path is obtained. The detour collision-free parking path with the smallest real-time risk value is then selected as the new collision-free parking path and executed.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic parking, and in particular to an automatic parking obstacle avoidance method based on vehicle-mounted image data. Background Art

[0002] In modern society, with the rapid increase in car ownership, parking has become an increasingly prominent problem. To improve parking convenience and safety, automated parking technology has emerged. It helps drivers automatically find a suitable parking space and control the vehicle to complete the parking maneuver, significantly reducing the driver's operational burden during the parking process.

[0003] However, existing automated parking technologies still face several challenges in practical application. For one thing, due to factors such as the camera's installation position, angle, and hardware specifications, image data captured by different cameras can exhibit perspective differences and distortion. This results in inaccurate stitched multi-view scene images, making it difficult to truly reflect the actual parking scenario, thus impacting the accuracy of subsequent parking path planning.

[0004] On the other hand, during parking, objects in the surrounding environment (such as other vehicles, pedestrians, and obstacles) change dynamically. Existing parking path planning methods often cannot effectively and timely respond to these dynamic changes. When encountering obstacles, it may be impossible to quickly plan a reasonable detour path, or the planned detour path may have a high risk of causing a collision. To this end, an automatic parking obstacle avoidance method based on on-board image data is proposed. Summary of the Invention

[0005] In order to solve the above technical problems, the purpose of the present invention is to provide an automatic parking obstacle avoidance method based on vehicle-mounted image data.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] The automatic parking obstacle avoidance method based on vehicle-mounted image data includes the following steps:

[0008] Step S1: Establish a spatial calibration plate, map parking scene image data captured by the on-board surround-view camera and the front-view fisheye camera onto the spatial calibration plate, mark several corner points on the plate, match the corner points of different parking scene image data, and mark some corner points as inliers based on the matching results. Then, adjust the hardware parameters of the on-board surround-view camera based on the positions of the inliers. Then, splice the parking scene image data captured by the on-board surround-view camera and the front-view fisheye camera after the adjustments to obtain a multi-view scene image.

[0009] Step S2: establishing a dynamic parking scene model centered on the vehicle based on the multi-view scene images, dividing the dynamic parking scene model into several scene object models, and setting scene semantic labels for each scene object model;

[0010] Step S3: Select a target parking position in the dynamic parking scene model, generate a corresponding collision-free parking path, set a scene model update cycle, and then obtain the spatial position of the scene object model in the next scene model update cycle;

[0011] Step S4: Determine whether a collision will occur on the collision-free parking path based on the spatial position of the scene object model in the next scene model update cycle. Generate multiple detour collision-free parking paths based on the judgment result, obtain the real-time risk value of each detour collision-free parking path, and then select the detour collision-free parking path with the smallest real-time risk value as the new collision-free parking path and execute it.

[0012] Furthermore, the process of marking corner points in the parking scene image data includes:

[0013] Four surround-view cameras and one forward-facing fisheye camera are deployed on the vehicle. The surround-view cameras are located in the front, rear, left, right, and forward wide-angle areas of the vehicle. The surround-view cameras and forward-facing fisheye cameras are spatially distributed, and the camera shooting ranges at adjacent spatial locations overlap.

[0014] When a vehicle generates a parking request, the system first uses the onboard surround-view cameras and a forward-looking fisheye camera to simultaneously capture 10 to 20 parking scene images from different angles and create a spatial calibration plate.

[0015] The spatial calibration plate is composed of a number of spatial grids of the same size, and the volume of each spatial grid is equal to a spatial cube composed of six pixel surfaces;

[0016] Each parking scene image data is mapped to the spatial calibration plate according to the camera and shooting angle, and several corner points are marked in each parking scene image data through OpenCV.

[0017] Furthermore, the process of matching corner points of different parking scene image data includes:

[0018] Using the parking scene image data captured by the front-view fisheye camera as reference image data, firstly, matching corner points in the parking scene image data captured by the front-view fisheye camera and the adjacent onboard surround-view camera;

[0019] If two corner points correspond to the same position, the two corner points are considered matched, otherwise they are considered unmatched, and the matching corner points are grouped, with each group containing several corner points.

[0020] Furthermore, the process of adjusting the hardware parameters of each vehicle-mounted surround view camera includes:

[0021] Obtaining model parameters for each group of corner points, wherein the model parameters are used to represent coordinate position parameters of corner points of different parking scene image data mapped onto another parking scene image data according to a matching relationship;

[0022] Set the planar distance threshold and spatial distance threshold, and use the model parameters to obtain the corner points corresponding to any on-board surround-view camera in the same group. Map them to the coordinate positions of the parking scene image data collected by the forward-looking fisheye camera, and then obtain the coordinate distance between the two corresponding corner points in the same parking scene image data.

[0023] If the coordinate distance is greater than or equal to the plane distance threshold, no action is taken;

[0024] If the coordinate distance is less than the plane distance threshold, the corresponding two corner points are recorded as inliers;

[0025] Then, based on the internal point correspondence between the image data of each parking scene, the Zhang Zhengyou calibration method is used to obtain the intrinsic parameter matrix and extrinsic parameter matrix of each camera;

[0026] Obtaining the world coordinate position of each interior point through the intrinsic parameter matrix and extrinsic parameter matrix of each camera, wherein the world coordinate position represents the actual spatial position of each interior point according to the hardware parameters of the camera;

[0027] The world coordinate points of the corresponding inner points of each on-board surround-view camera are mapped to the spatial calibration plate, and the spatial distance between the coordinate position of each inner point on the spatial calibration plate and the corresponding world coordinate point is determined. If the spatial distance is greater than or equal to the spatial distance threshold, the hardware parameters of the corresponding on-board surround-view camera are adjusted until the spatial distance is less than the spatial distance threshold.

[0028] Furthermore, the process of acquiring the multi-view scene image includes:

[0029] After the hardware parameters of the surround-view cameras adjacent to the front-view fisheye camera are adjusted, the same method is used to adjust the hardware parameters of the remaining two surround-view cameras based on the parking scene image data collected by the surround-view cameras adjacent to the front-view fisheye camera.

[0030] After all the on-board surround-view cameras have completed hardware parameter adjustment, each camera continues to collect parking scene image data within its shooting range. Based on the internal point correspondence between the on-board surround-view cameras and the front-view fisheye camera, the parking scene image data collected at the same time node are spliced ​​to obtain a multi-view scene image.

[0031] Furthermore, the process of dividing a plurality of scene object models in the dynamic parking scene model includes:

[0032] A dynamic parking scene model of the parking scene is established based on the multi-view scene images, and a three-dimensional coordinate system is established with the vehicle as the origin, and the dynamic parking scene model is then mapped into the three-dimensional coordinate system;

[0033] Divide the dynamic parking scene model into a number of model voxels of equal size, wherein the model voxels are in the shape of cubes, and each face has an independent plane pixel value;

[0034] Starting from the model voxels at the eight corner positions of the dynamic parking scene model, pixel averaging is performed on each model voxel with its adjacent model voxels. That is, the pixels on the same side of the model voxel and the adjacent model voxels are added together and the average value is taken. The average value is then reassigned to the corresponding side. No operation is performed on the positions where the model voxel has no adjacent faces.

[0035] Repeat the above model voxel pixel averaging operation until the model voxels in eight directions converge at the same position for model voxel pixel averaging;

[0036] After the dynamic parking scene model completes voxel averaging, the voxel pixels are clustered using the Gaussian mixture clustering algorithm. Several scene object models are then divided in the dynamic parking scene model, and scene semantic labels are set for each scene object model.

[0037] Set a scene model update cycle, randomly calibrate n motion feature points for each scene object model, and then detect whether each motion feature point has moved every time a scene model update cycle ends, where n is a natural number greater than 0;

[0038] If the motion feature point is detected to be displaced in three consecutive scene model update cycles, the corresponding scene object model is recorded as a dynamic object scene model, otherwise it is recorded as a static scene object model.

[0039] Furthermore, the process of obtaining the spatial position of the scene object model in the next scene model update cycle includes:

[0040] The target parking location is selected in the dynamic parking scene model. Then, an improved rapid exploration random tree algorithm is used to mark the vehicle's current location as the starting point and the target parking location as the end point in the dynamic parking scene model. Various dynamic object scene models and static scene object models are represented as obstacles. A collision-free parking path is traversed in the dynamic parking scene model.

[0041] At the beginning of each scene model update cycle, the prediction uncertainty is quantified based on Gaussian process regression to predict the spatial position of each dynamic object scene model in the next scene model update cycle;

[0042] The predicted spatial coordinates of the motion feature points of each dynamic object scene model are mapped to the displacement position of each dynamic object scene model in the dynamic parking scene model, and the displacement direction is marked.

[0043] Furthermore, the process of generating the detour collision-free parking path includes:

[0044] Whenever the dynamic parking scene model updates the displacement position of each dynamic object scene model in the next scene model update cycle, determining whether the vehicle collides along the collision-free parking path in the next scene model update cycle;

[0045] If it is determined that no collision occurs, no action is taken;

[0046] If a collision is determined to have occurred, multiple detour collision-free parking paths are regenerated with the vehicle's current position as the starting point and the target parking position as the end point. Each detour collision-free parking path is divided into several driving path nodes based on the vehicle's current driving speed and the length of the scenario model update cycle, and the real-time risk value of each detour collision-free parking path is obtained.

[0047] Furthermore, the process of obtaining the real-time risk value of each detour collision-free parking path includes:

[0048] A real-time risk threshold is set and the real-time risk value of the detour collision-free parking path with the minimum real-time risk value is compared. If the minimum real-time risk value is less than or equal to the real-time risk threshold, the corresponding detour collision-free parking path replaces the original dynamic collision-free parking path.

[0049] If the minimum real-time risk value is greater than the real-time risk threshold, then no parking path is currently executable;

[0050] The above determination process is repeated every time a scene model update cycle passes until the vehicle reaches the target parking position.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. This invention maps parking scene image data collected by the on-board surround-view camera and the front-view fisheye camera onto a spatial calibration plate, annotates corner points for matching, and adjusts the hardware parameters of the on-board surround-view camera based on the matching results. This effectively eliminates the perspective differences and distortion between images collected by different cameras, making the spliced ​​multi-perspective scene images more accurately reflect the actual parking scene, providing a more reliable data foundation for subsequent parking path planning.

[0053] 2. Based on the spatial position of the scene object model in the next scene model update cycle, it is determined whether the collision-free parking path will occur, and multiple detour collision-free parking paths are generated. The detour collision-free parking path with the smallest real-time risk value is selected as the new execution path. This realizes dynamic planning of the parking path and can promptly respond to dynamic changes in the surrounding environment during the parking process, effectively avoiding the occurrence of collision accidents and improving the safety and reliability of automatic parking. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention.

[0055] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0056] To make the objectives, technical solutions, and advantages of the present invention more apparent, the technical solutions of the present invention will be described in detail below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other implementations obtained by those of ordinary skill in the art without inventive effort are within the scope of protection of the present invention.

[0057] like Figure 1 As shown, the automatic parking obstacle avoidance method based on vehicle-mounted image data includes the following steps:

[0058] Step S1: Establish a spatial calibration plate, map parking scene image data captured by the on-board surround-view camera and the front-view fisheye camera onto the spatial calibration plate, mark several corner points on the plate, match the corner points of different parking scene image data, and mark some corner points as inliers based on the matching results. Then, adjust the hardware parameters of the on-board surround-view camera based on the positions of the inliers. Then, splice the parking scene image data captured by the on-board surround-view camera and the front-view fisheye camera after the adjustments to obtain a multi-view scene image.

[0059] Step S2: establishing a dynamic parking scene model centered on the vehicle based on the multi-view scene images, dividing the dynamic parking scene model into several scene object models, and setting scene semantic labels for each scene object model;

[0060] Step S3: Select a target parking position in the dynamic parking scene model, generate a corresponding collision-free parking path, set a scene model update cycle, and then obtain the spatial position of the scene object model in the next scene model update cycle;

[0061] Step S4: Determine whether a collision will occur on the collision-free parking path based on the spatial position of the scene object model in the next scene model update cycle. Generate multiple detour collision-free parking paths based on the judgment result, obtain the real-time risk value of each detour collision-free parking path, and then select the detour collision-free parking path with the smallest real-time risk value as the new collision-free parking path and execute it.

[0062] Furthermore, step S1 is implemented by the following process:

[0063] Four surround-view cameras and one forward-facing fisheye camera are deployed on the vehicle. The surround-view cameras are located in the front, rear, left, right, and forward wide-angle areas of the vehicle. The surround-view cameras and forward-facing fisheye cameras are spatially distributed, and the camera shooting ranges at adjacent spatial locations overlap.

[0064] When a vehicle generates a parking request, the system first uses the onboard surround-view cameras and a forward-looking fisheye camera to simultaneously capture 10 to 20 parking scene images from different angles and create a spatial calibration plate.

[0065] The spatial calibration plate is composed of a number of spatial grids of the same size, and the volume of each spatial grid is equal to a spatial cube composed of six pixel surfaces;

[0066] Map each parking scene image data onto a spatial calibration plate according to the camera and shooting angle, and use OpenCV to mark several corner points in each parking scene image data;

[0067] Since the shooting ranges of adjacent cameras overlap, the parking scene image data captured by adjacent cameras must contain identical portions. Therefore, using the parking scene image data captured by the forward-looking fisheye camera as reference image data, the corner points within the parking scene image data captured by the forward-looking fisheye camera and its adjacent onboard surround-view camera are first matched.

[0068] If two corner points are determined to correspond to the same position, they are considered to match; otherwise, they are considered to not match. Furthermore, since each onboard surround-view camera and front-view fisheye camera captures multiple parking scene image data from different angles, the matching corner points can be grouped, with each group containing several corner points.

[0069] Obtaining model parameters for each group of corner points, wherein the model parameters are used to represent coordinate position parameters of corner points of different parking scene image data mapped onto another parking scene image data according to a matching relationship;

[0070] Set the planar distance threshold and spatial distance threshold, and use the model parameters to obtain the corner points corresponding to any on-board surround-view camera in the same group. Map them to the coordinate positions of the parking scene image data collected by the forward-looking fisheye camera, and then obtain the coordinate distance between the two corresponding corner points in the same parking scene image data.

[0071] If the coordinate distance is greater than or equal to the plane distance threshold, no action is taken;

[0072] If the coordinate distance is less than the plane distance threshold, the corresponding two corner points are recorded as inliers;

[0073] Then, according to the internal point correspondence between the image data of each parking scene, the internal parameter matrix of each camera (focal length f x , f y , principal point offset c x , c y ) and external parameter matrix (rotation matrix R, translation vector T);

[0074] Obtaining the world coordinate position of each interior point through the intrinsic parameter matrix and extrinsic parameter matrix of each camera, wherein the world coordinate position represents the actual spatial position of each interior point according to the hardware parameters of the camera;

[0075] The world coordinate points of the corresponding inner points of each on-board surround-view camera are mapped to the spatial calibration plate, and the spatial distance between the coordinate position of each inner point on the spatial calibration plate and the corresponding world coordinate point is determined. If the spatial distance is greater than or equal to the spatial distance threshold, the hardware parameters of the corresponding on-board surround-view camera are adjusted until the spatial distance is less than the spatial distance threshold.

[0076] Furthermore, after the hardware parameters of the surround-view cameras adjacent to the front-view fisheye camera are adjusted, the same method is used to adjust the hardware parameters of the remaining two surround-view cameras based on the parking scene image data collected by the surround-view cameras adjacent to the front-view fisheye camera.

[0077] After all the on-board surround-view cameras have completed hardware parameter adjustment, each camera continues to collect parking scene image data within its shooting range. Based on the internal point correspondence between the on-board surround-view cameras and the front-view fisheye camera, the parking scene image data collected at the same time node are spliced ​​to obtain a multi-view scene image.

[0078] Furthermore, step S2 is implemented by the following process:

[0079] A dynamic parking scene model of the parking scene is established based on the multi-view scene images, and a three-dimensional coordinate system is established with the vehicle as the origin, and the dynamic parking scene model is then mapped into the three-dimensional coordinate system;

[0080] Divide the dynamic parking scene model into a number of model voxels of equal size, wherein the model voxels are in the shape of cubes, and each face has an independent plane pixel value;

[0081] Starting from the model voxels at the eight corner positions of the dynamic parking scene model, pixel averaging is performed on each model voxel with its adjacent model voxels. That is, the pixels on the same side of the model voxel and the adjacent model voxels are added together and the average value is taken. The average value is then reassigned to the corresponding side. No operation is performed on the positions where the model voxel has no adjacent faces.

[0082] Repeat the above model voxel pixel averaging operation until the model voxels in eight directions converge at the same position for model voxel pixel averaging;

[0083] Among them, the eight vertex positions are front, back, left, right, left front, left back, right front and right back;

[0084] After the dynamic parking scene model completes voxel averaging, each voxel is clustered using a Gaussian mixture clustering algorithm. This divides the dynamic parking scene model into several scene object models, and assigns scene semantic labels to each scene object model. These labels include pedestrians, vehicles, cones, curbs, parking spaces, and more.

[0085] Set a scene model update cycle, the frequency and duration of which is generally 10ms. Randomly calibrate n motion feature points for each scene object model. Then, at the end of each scene model update cycle, detect whether each motion feature point has moved. n is a natural number greater than 0.

[0086] If the motion feature point is detected to be displaced in three consecutive scene model update cycles, the corresponding scene object model is recorded as a dynamic object scene model, otherwise it is recorded as a static scene object model.

[0087] Furthermore, step S3 is implemented by the following process:

[0088] The user selects a target parking location in the dynamic parking scene model based on semantic labels. Then, using an improved rapid exploration random tree algorithm, the vehicle's current location is labeled as the starting point and the target parking location as the end point. Dynamic and static scene object models are used as obstacles to traverse a collision-free parking path within the dynamic parking scene model.

[0089] At the beginning of each scene model update cycle, the prediction uncertainty is quantified based on Gaussian process regression to predict the spatial position of each dynamic object scene model in the next scene model update cycle. The formula is:

[0090] ;

[0091] in represents the predicted spatial coordinates of the motion feature points of the dynamic object scene model in the dynamic parking scene model during the t-th scene model update cycle, 、 、 Represents the predicted position mean of the three-dimensional Gaussian distribution in the x-axis, y-axis, and z-axis directions, Indicates the standard deviation value of the predicted position in each direction;

[0092] The predicted spatial coordinates of the motion feature points of each dynamic object scene model are mapped to the displacement position of each dynamic object scene model in the dynamic parking scene model, and the displacement direction is marked.

[0093] Furthermore, step S4 is implemented by the following process:

[0094] Whenever the dynamic parking scene model updates the displacement position of each dynamic object scene model in the next scene model update cycle, it is determined whether the vehicle will collide with the next scene model update cycle along the collision-free parking path;

[0095] If it is determined that no collision occurs, no action is taken;

[0096] If a collision is detected, multiple collision-free parking paths are regenerated with the vehicle's current position as the starting point and the target parking position as the end point. Each collision-free parking path is then divided into several nodes based on the vehicle's current speed and the duration of the scenario model update cycle.

[0097] And calculate the real-time risk value of each detour collision-free parking path;

[0098] The calculation formula of the real-time risk value is:

[0099] ;

[0100] ;

[0101] ;

[0102] ;

[0103] in, represents the real-time risk value of each detour collision-free parking path, 、 、 represents the weight parameter, Indicates the risk influencing factors corresponding to the motion feature points, Indicates the risk influencing factors corresponding to the parking path length, represents the risk influencing factor corresponding to the steering angle change rate, 、 、 、 、 、 represents the coordinate value of the ith driving path node and the coordinate value of the nearest motion feature point on its nearest dynamic object scene model. L represents the total length of the corresponding detour collision-free parking path. represents the steering angle change rate of the i-th driving path node, and N represents the total number of driving path nodes;

[0104] A real-time risk threshold is set, and the real-time risk value of the detour collision-free parking path with the minimum real-time risk value is compared with the real-time risk threshold. If the minimum real-time risk value is less than or equal to the real-time risk threshold, the corresponding detour collision-free parking path replaces the original dynamic collision-free parking path.

[0105] If the minimum real-time risk value is greater than the real-time risk threshold, then no parking path is currently executable;

[0106] The above determination process is repeated every time a scene model update cycle passes until the vehicle reaches the target parking position.

[0107] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An automatic parking obstacle avoidance method based on vehicle-mounted image data, characterized in that: The following steps are involved: Step S1: Establish a spatial calibration plate, map parking scene image data captured by the on-board surround-view camera and the front-view fisheye camera onto the spatial calibration plate, mark several corner points on the plate, match the corner points of different parking scene image data, and mark some corner points as inliers based on the matching results. Then, adjust the hardware parameters of the on-board surround-view camera based on the positions of the inliers. Then, splice the parking scene image data captured by the on-board surround-view camera and the front-view fisheye camera after the adjustments to obtain a multi-view scene image. Step S2: establishing a dynamic parking scene model centered on the vehicle based on the multi-view scene images, dividing the dynamic parking scene model into several scene object models, and setting scene semantic labels for each scene object model; Step S3: Select a target parking position in the dynamic parking scene model, generate a corresponding collision-free parking path, set a scene model update cycle, and then obtain the spatial position of the scene object model in the next scene model update cycle; Step S4: Determine whether a collision will occur on the collision-free parking path based on the spatial position of the scene object model in the next scene model update cycle. Generate multiple detour collision-free parking paths based on the determination result, obtain a real-time risk value for each detour collision-free parking path, and then select the detour collision-free parking path with the lowest real-time risk value as the new collision-free parking path and execute it. The process of matching corner points of different parking scene image data includes: Using the parking scene image data captured by the front-view fisheye camera as reference image data, matching corner points in the parking scene image data captured by the front-view fisheye camera and the adjacent on-board surround-view camera; If two corner points correspond to the same position, the two corner points are considered matched, otherwise they are considered unmatched, and the matching corner points are grouped, with each group containing several corner points.

2. The automatic parking obstacle avoidance method based on vehicle-mounted image data according to claim 1, characterized in that: The process of marking corner points in parking scene image data includes: Four surround-view cameras and one front-view fisheye camera are deployed on the vehicle. The surround-view cameras and the front-view fisheye camera are distributed according to their spatial positions, and the camera shooting ranges at adjacent spatial positions overlap. When a vehicle generates a parking request, each onboard surround-view camera and a forward-looking fisheye camera simultaneously capture parking scene image data from different angles and create a spatial calibration plate; Each parking scene image data is mapped to the spatial calibration plate according to the camera and shooting angle, and several corner points are marked in each parking scene image data through OpenCV.

3. The automatic parking obstacle avoidance method based on vehicle-mounted image data according to claim 2, characterized in that: The process of adjusting the hardware parameters of each vehicle surround view camera includes: Obtaining model parameters for each group of corner points, wherein the model parameters are used to represent coordinate position parameters of corner points of different parking scene image data mapped onto another parking scene image data according to a matching relationship; Set the planar distance threshold and spatial distance threshold, obtain the corner points corresponding to any on-board surround-view camera in the same group based on the model parameters, map them to the coordinate positions of the parking scene image data collected by the forward-looking fisheye camera, and then obtain the coordinate distance between the two corresponding corner points in the same parking scene image data; If the coordinate distance is greater than or equal to the plane distance threshold, no operation is performed. If the coordinate distance is less than the plane distance threshold, the corresponding two corner points are recorded as inliers. Then, the intrinsic and extrinsic parameter matrices of each camera are obtained based on the inlier point correspondence between the parking scene image data. The world coordinate position of each internal point is obtained through the intrinsic parameter matrix and extrinsic parameter matrix of each camera, and the world coordinate position of the internal point corresponding to each vehicle-mounted surround-view camera is mapped to the spatial calibration plate. The spatial distance between the coordinate position of each internal point on the spatial calibration plate and the corresponding world coordinate position is determined. If the spatial distance is greater than or equal to the spatial distance threshold, the hardware parameters of the corresponding vehicle-mounted surround-view camera are adjusted until the spatial distance is less than the spatial distance threshold.

4. The automatic parking obstacle avoidance method based on vehicle-mounted image data according to claim 3, characterized in that: The process of acquiring the multi-view scene image includes: After all the on-board surround-view cameras have completed hardware parameter adjustment, each camera continues to collect parking scene image data within its shooting range. Based on the internal point correspondence between the on-board surround-view cameras and the front-view fisheye camera, the parking scene image data collected at the same time node are spliced ​​to obtain a multi-view scene image.

5. The automatic parking obstacle avoidance method based on vehicle-mounted image data according to claim 4, characterized in that: The process of dividing several scene object models in the dynamic parking scene model includes: A dynamic parking scene model of the parking scene is established based on the multi-view scene images, and a three-dimensional coordinate system is established with the vehicle as the origin, and the dynamic parking scene model is then mapped into the three-dimensional coordinate system; The dynamic parking scene model is divided into a number of model voxels of the same size. Starting from the model voxels at the eight corner positions of the dynamic parking scene model, pixel averaging is performed on each model voxel and its adjacent model voxels. After the dynamic parking scene model completes voxel averaging, the voxel pixels are clustered using the Gaussian mixture clustering algorithm. Several scene object models are then divided in the dynamic parking scene model, and scene semantic labels are set for each scene object model. Set a scene model update cycle, randomly calibrate n motion feature points for each scene object model, and then detect whether each motion feature point has moved every time a scene model update cycle ends, where n is a natural number greater than 0; If the motion feature point is detected to be displaced in three consecutive scene model update cycles, the corresponding scene object model is recorded as a dynamic object scene model, otherwise it is recorded as a static scene object model.

6. The automatic parking obstacle avoidance method based on vehicle-mounted image data according to claim 5, characterized in that: The process of obtaining the spatial position of the scene object model in the next scene model update cycle includes: A target parking position is selected in a dynamic parking scene model, the current position of the vehicle is marked as a starting point, the target parking position is marked as an end point, and each scene object model is marked as an obstacle, and a collision-free parking path is traversed in the dynamic parking scene model; At the start of each scene model update cycle, predict the spatial position of each dynamic object scene model in the next scene model update cycle; The predicted spatial coordinates of the motion feature points of each dynamic object scene model are mapped to the displacement positions of each dynamic object scene model in the dynamic parking scene model.

7. The automatic parking obstacle avoidance method based on vehicle-mounted image data according to claim 6, characterized in that: The process of generating the detour collision-free parking path includes: Whenever the dynamic parking scene model updates the displacement position of each dynamic object scene model in the next scene model update cycle, determining whether the vehicle collides along the collision-free parking path in the next scene model update cycle; Based on the judgment results, multiple detour collision-free parking paths are regenerated, and each detour collision-free parking path is divided into several driving path nodes based on the vehicle's current driving speed and the duration of the scenario model update cycle, and the real-time risk value of each detour collision-free parking path is obtained.

8. The automatic parking obstacle avoidance method based on vehicle-mounted image data according to claim 7, characterized in that: The process of obtaining the real-time risk value of each detour collision-free parking path includes: A real-time risk threshold is set and the real-time risk value of the detour collision-free parking path with the minimum real-time risk value is compared. If the minimum real-time risk value is less than or equal to the real-time risk threshold, the corresponding detour collision-free parking path replaces the original dynamic collision-free parking path. If the minimum real-time risk value is greater than the real-time risk threshold, then no parking path is currently executable.

Images