Obstacle contour perception method based on grating grid in parking scenario

Abstract

This invention provides a method for obstacle contour perception based on grating mesh in parking scenarios. A grating mesh is generated by a point light source and grating lenses and projected onto the area to be measured. Video data of the area to be measured is acquired using a monocular camera. The video data is preprocessed to obtain a grating map of the region of interest (ROI) of the obstacle. Grating picking is performed on the grating map of the ROI to extract grating mesh lines, and these lines are numbered. Feature points are then picked up, and 3D feature point reconstruction is completed. Based on the reconstructed 3D feature points, a normal-multiple clustering method is used to fit the facade, outlining the obstacle contour and achieving obstacle contour perception. Compared with existing technologies, this method can effectively improve the speed and accuracy of obstacle contour visual detection during parking, and enhance the adaptability of visual detection to the environment. It has the advantages of low cost, high accuracy, and ease of use.

Description

Technical Field

[0001] This invention relates to a method for obstacle contour perception based on grating mesh in parking scenarios, belonging to the field of object recognition technology. Background Technology

[0002] Real-time obstacle perception is essential when a vehicle is parking. Typically, obstacle information is simplified into a two-dimensional representation, with the sensor-detected obstacle information serving as the boundary of the drivable area. Within this area, a safe parking process can be completed. Therefore, effective obstacle perception helps vehicles park safely and efficiently.

[0003] Currently, obstacle perception in parking scenarios is mainly accomplished through various types of radar or vision, with most automatic parking systems using a combination of radar and cameras.

[0004] Existing ultrasonic radar can provide distance information between surrounding objects and the vehicle itself, enabling obstacle detection, but with low accuracy; binocular cameras can provide information about surrounding objects through image recognition algorithms, but suffer from a significant drop in accuracy for distant perception, and visual detection is not suitable for object perception in dim environments; lidar can provide three-dimensional information, but is very expensive.

[0005] For example, the depth contour estimation method based on a binocular RGB-D camera disclosed in Chinese invention patent ZL201711311829.3 also suffers from low accuracy in long-distance obstacle perception and high cost.

[0006] The above-mentioned issues are problems that should be considered and solved in the process of obstacle contour perception based on grating mesh in parking scenarios. Summary of the Invention

[0007] The purpose of this invention is to provide an obstacle contour perception method based on grating mesh in parking scenarios, which solves the problems of low accuracy in long-distance obstacle perception during parking and the need to improve the adaptability of visual detection to the environment in the existing technology.

[0008] The technical solution of this invention is:

[0009] A method for obstacle contour perception based on raster mesh in parking scenarios includes the following steps:

[0010] S1. A grating grid is generated by a point light source and a grating lens and projected onto the area to be measured. Video data of the area to be measured is acquired by a monocular camera. The video data is preprocessed to obtain a grating map of the region of interest (ROI) of the obstacle.

[0011] S2. For the raster map of the Region of Interest (ROI) of the pre-obstacle, perform raster pickup, extract the raster grid lines and grid points, and number the grid lines;

[0012] S3. Perform feature point picking, including feature point extraction and normalization processing;

[0013] S4. Based on the planar pixel coordinates of the feature points obtained in step S4, perform a conditional search in the pre-calibrated database to realize the transformation of planar pixel points to spatial three-dimensional coordinates and complete the reconstruction of three-dimensional feature points.

[0014] S5. Based on the 3D feature points reconstructed in step S4, the facade is fitted using the normal multiple clustering method to outline the obstacle and realize obstacle outline perception.

[0015] Furthermore, in step S1, the video data preprocessing includes grayscale conversion, binarization, image enhancement, and picking of the region of interest (ROI).

[0016] Further, in step S2, the raster image of the region of interest (ROI) of the pre-obstacle is raster-picked to extract the raster grid lines and grid points, and the grid lines are numbered. Specifically,

[0017] S21. Use a line segment detection algorithm to extract the raster lines from the raster map of the region of interest (ROI) of the pre-obstacle;

[0018] S22. Perform grating line fusion, including the process of classification-splitting-fitting, and then number the fitted grating lines again in the horizontal and vertical directions.

[0019] Further, in step S22, the grating lines are fused, including the process of classification-splitting-fitting. Specifically, firstly, all grating lines are classified horizontally and vertically based on the direction of the grating lines. Then, all grating lines of the same type are split into scattered points with a step size l. Finally, the least squares method is used to fit the split scattered points to obtain the fitted grating lines.

[0020] Further, in step S3, feature point picking is performed, including feature point extraction and normalization processing, specifically as follows:

[0021] S31. Extract feature points. Extract all intersections of horizontal and vertical lines from step S2 according to the line number, which are the feature points. Number the intersections according to the horizontal and vertical lines they belong to.

[0022] S32. Perform feature point normalization processing, including point density processing and point missing processing.

[0023] Further, in step S32, the point density processing specifically involves using the density clustering method DBSCAN to cluster points in regions where the point distribution density is higher than a set value. At the same time, a KD tree, i.e., a k-dimensional tree, is used to limit the number of points in the region. The points in the region are divided into multiple classes, and each class is assigned to multiple corresponding center points to replace the current class. The center points of the newly acquired replacement classes are numbered using the smallest point number in the dense point set, until the processing of all regions where the point distribution density is higher than the set value is completed.

[0024] Furthermore, in step S32, the point missing handling specifically involves:

[0025] S321. Record the points obtained after point densification and the remaining points extracted in step S31 that have not undergone point densification on the new image. Then, traverse the points of the new image from left to right and from top to bottom to determine the adjacent points of each point in the four directions of up, down, left, and right and calculate the adjacent distances.

[0026] S322. Set a threshold D. When the adjacent distance d0 of one of the four directions (up, down, left, right) is greater than D, it is considered that there is a missing point in that direction. Find points within a range of 2D near the set area and record the number of points n. Set a threshold N. When n < N, expand the search range until n ≥ N. Then, connect the points horizontally and vertically according to their numbers and extract the intersection of the lines as the supplementary points for the missing positions. Otherwise, directly connect the points horizontally and vertically according to their numbers and extract the intersection as the supplementary points.

[0027] The beneficial effects of this invention are:

[0028] I. This obstacle contour perception method based on grating mesh in parking scenarios, compared with existing technologies, can effectively improve the accuracy of obstacle visual detection of obstacle contours during parking and enhance the adaptability of visual detection to the environment. It has the advantages of low cost, high accuracy and ease of use.

[0029] II. This invention enhances the scene of the test area by projecting a grating grid onto the test area, improving the situation where visual perception is impossible due to dim lighting. Furthermore, during visual perception, the invention uses the refraction generated when the grid lines are projected onto the obstacle to calculate the existence and spatial coordinates of the obstacle, thereby improving the accuracy of long-distance obstacle contour perception during parking.

[0030] Third, the obstacle contour perception method based on grating mesh in this parking scenario uses only a monocular camera in terms of hardware, which reduces hardware costs and the complexity of subsequent deployment. Attached Figure Description

[0031] Figure 1This is a flowchart illustrating the obstacle contour perception method based on grating mesh in a parking scenario according to an embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram comparing the three-dimensional spatial coordinate values ​​obtained in the embodiment with the actual values. Detailed Implementation

[0033] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0034] Example

[0035] A method for obstacle contour perception based on raster mesh in parking scenarios, such as... Figure 1 This includes the following steps:

[0036] S1. A grating grid is generated by a point light source and a grating lens and projected onto the area to be measured. Video data of the area to be measured is acquired by a monocular camera. The video data is preprocessed to obtain a grating map of the region of interest (ROI) of the obstacle.

[0037] In step S1, the video data is preprocessed, including grayscale conversion, binarization, image enhancement, and region of interest (ROI) selection.

[0038] S2. For the raster map of the Region of Interest (ROI) of the pre-obstacle, perform raster picking, extract the raster grid lines, and number the grid lines.

[0039] S21. Use a line segment detection algorithm to extract the raster lines from the raster map of the region of interest (ROI) of the pre-obstacle;

[0040] S22. Perform grating line fusion, including the process of classification-splitting-fitting, and then number the fitted grating lines again in the horizontal and vertical directions.

[0041] In step S22, the grating lines are fused, including a classification-splitting-fitting process. Specifically, firstly, all grating lines are classified horizontally and vertically based on their orientation. Then, grating lines of the same category are split into scattered points with a step size l. Finally, the least squares method is used to fit the split scattered points to obtain the fitted grating lines. This solves the problems of insufficient line length, partial overlap, and discontinuity in the extracted lines.

[0042] S3. Perform feature point picking, including feature point extraction and normalization.

[0043] S31. Extract the points. Extract all intersections of horizontal and vertical lines in step S2, i.e., feature points, according to the line number, and number the intersections according to the horizontal and vertical lines they belong to.

[0044] S32. Perform point normalization processing, including point density processing and point missing processing.

[0045] In step S32, the point density processing specifically involves using the density clustering method DBSCAN to cluster points in regions where the point distribution density is higher than a set value. At the same time, a KD tree (k-dimensional tree) is used to limit the number of points, dividing the points in the region into multiple classes. Each class is then assigned a corresponding set of multiple center points to replace the current class. The center points of the newly acquired replacement classes are numbered using the smallest point number in the dense point set, until the processing of all regions where the point distribution density is higher than the set value is completed.

[0046] In step S32, point-dense processing effectively reduces noise in the image, normalizes data of the same type, facilitates further processing of the data, and ensures the running speed of the program.

[0047] In step S32, the missing point handling specifically involves:

[0048] S321. Record the points obtained after point densification and the remaining points extracted in step S31 that have not undergone point densification on the new image. Then, traverse the points of the new image from left to right and from top to bottom to determine the adjacent points of each point in the four directions of up, down, left, and right and calculate the adjacent distances.

[0049] S322. Set a threshold D. When the adjacent distance d0 of one of the four directions (up, down, left, right) is greater than D, it is considered that there is a missing point in that direction. Find points within a range of 2D near the set area and record the number of points n. Set a threshold N. When n < N, expand the search range until n ≥ N. Then, connect the points horizontally and vertically according to their numbers and extract the intersection of the lines as the supplementary points for the missing positions. Otherwise, directly connect the points horizontally and vertically according to their numbers and extract the intersection as the supplementary points.

[0050] S4. Based on the planar pixel coordinates of the feature points obtained in step S4, perform a conditional search in the pre-calibrated database to realize the transformation of planar pixel points to spatial three-dimensional coordinates and complete the reconstruction of three-dimensional feature points.

[0051] S5. Based on the 3D feature points reconstructed in step S4, the facade is fitted using the normal multiple clustering method to outline the obstacle and realize obstacle outline perception.

[0052] This obstacle contour perception method based on grating mesh in parking scenarios effectively improves the speed and accuracy of obstacle visual detection during parking compared to existing technologies, and enhances the adaptability of visual detection to the environment. It boasts advantages such as low cost, high accuracy, and ease of use.

[0053] This invention enhances the scene of the test area by projecting a grating grid onto it, improving the situation where visual perception is impossible due to dim lighting. Furthermore, during visual perception, the invention uses the refraction generated when the grid lines illuminate obstacles to calculate the presence and spatial coordinates of obstacles, thereby improving the accuracy of long-distance obstacle contour perception during parking.

[0054] This obstacle contour perception method based on grating mesh in parking scenarios generates the grating mesh through point light sources and grating lenses, and acquires it through an onboard camera. The object to be measured is placed in the projected grating mesh area, and the onboard camera acquires images of the grating mesh area. The images are processed to extract the grating mesh lines and mesh points, and the mesh lines are numbered. The relative positional relationships between the points are used to find the corresponding spatial point positions, thereby outlining the contour of the obstacle.

[0055] The obstacle contour perception method based on grating mesh in this parking scenario of the embodiment is experimentally verified as follows:

[0056] Set up a test platform and fix the camera and grating projector. Connect the power supply to create a 3m x 7m grating projection area on the ground, which is the area to be tested. Lay a printed mesh cloth flat on the area to be tested (the horizontal and vertical lines on the mesh cloth are spaced 1cm apart) as a ruler to facilitate the acquisition of the object's true position data. Place a 6cm single cube object at any position in the test area. Read the true values ​​of the spatial position coordinates of the object's key corner points. Obtain the three-dimensional spatial coordinate values ​​output by the method of this embodiment. Compare the results of the two values, where some data are as follows: Figure 2 The error percentage is analyzed as shown in Table 1:

[0057] Table 1 shows the error percentage of the test results for the methods in the embodiments.

[0058]

[0059] Depend on Figure 2 The test results in Table 1 show that: (1) the detection rate of the method in this embodiment for objects smaller than 6cm is 100%. (2) the corner error distribution of the method in this embodiment is almost concentrated within 3cm, and the probability of the corner error of the obstacle being less than or equal to 1cm reaches 66%. (3) the high-precision sensing depth distance of the method in this embodiment reaches 6.9 meters. The above experimental results verify that the obstacle contour sensing method based on grating mesh in this parking scenario of the embodiment can perform high-precision sensing of obstacle contours.

[0060] This obstacle contour perception method based on grating mesh in parking scenarios acquires video data of the target area using a monocular camera. After preprocessing, feature points are processed based on the characteristics of the grating. Then, a normal-based multiple clustering method is used to construct the obstacle's facade, thereby outlining the obstacle and obtaining the 3D coordinates of its corner points. This enables high-precision, long-distance obstacle contour perception. Hardware-wise, it uses only a monocular camera, reducing hardware costs and the complexity of subsequent deployment.

[0061] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for obstacle contour perception based on grating mesh in a parking scenario, characterized in that: Includes the following steps, S1. A grating grid is generated by a point light source and a grating lens and projected onto the area to be tested. Video data of the area to be tested is acquired by a monocular camera. The video data is preprocessed to obtain a grating map of the region of interest (ROI) of the obstacle. In step S1, the preprocessing of the video data includes grayscale conversion, binarization, image enhancement, and picking the region of interest (ROI). S2. For the raster image of the Region of Interest (ROI) of the pre-obstacle, raster picking is performed to extract the raster grid lines and number the grid lines; specifically, S21. The raster lines of the raster image of the ROI of the pre-obstacle are extracted using a line segment detection algorithm; S22. The raster lines are fused, including the process of classification-splitting-fitting, and the fitted raster lines are numbered again in the horizontal and vertical directions respectively. S3. Perform feature point picking, including feature point extraction and normalization; specifically, S31. Extract feature points by extracting all intersections of horizontal and vertical lines from step S2 according to line number, and number the intersections according to the horizontal and vertical lines they belong to; S32. Perform feature point normalization, including point density processing and point missing processing. S4. Based on the planar pixel coordinates of the feature points obtained in step S3, perform a conditional search in the pre-calibrated database to realize the transformation of planar pixel points to spatial three-dimensional coordinates and complete the reconstruction of three-dimensional feature points. S5. Based on the 3D feature points reconstructed in step S4, the facade is fitted using the normal multiple clustering method to outline the obstacle and realize obstacle outline perception.

2. The obstacle contour perception method based on grating mesh in parking scenarios as described in claim 1, characterized in that: In step S22, the grating lines are fused, including a classification-splitting-fitting process. Specifically, firstly, all grating lines are classified horizontally and vertically based on their orientation. Then, grating lines of the same type are fused using a step size. All grating lines are broken down into scattered points, and finally, the least squares method is used to fit the broken scattered points to obtain the fitted grating lines.

3. The obstacle contour perception method based on grating mesh in parking scenarios as described in claim 1, characterized in that: In step S32, the point density processing specifically involves using the density clustering method DBSCAN to cluster points in regions where the point distribution density is higher than a set value. At the same time, a KD tree (k-dimensional tree) is used to limit the number of points in the region, dividing the points in the region into multiple classes. Each class is then assigned a corresponding set of multiple center points to replace the current class. The center points of the newly acquired replacement classes are numbered using the smallest point number in the dense point set, until the processing of all regions where the point distribution density is higher than the set value is completed.

4. The obstacle contour perception method based on grating mesh in parking scenarios as described in claim 1, characterized in that: In step S32, the missing point handling specifically involves: S321. Record the points obtained after point densification and the remaining points extracted in step S31 that have not undergone point densification on the new image. Then, traverse the points of the new image from left to right and from top to bottom to determine the adjacent points of each point in the four directions of up, down, left, and right and calculate the adjacent distances. S322, Set threshold When the adjacent distance in one of the four directions (up, down, left, right) is... If this occurs, it is assumed that a point is missing in that direction, and a search is conducted in the vicinity of the specified area. Record the number of points within the range. Set threshold ,when Then expand the search range until... Then, connect the points horizontally and vertically according to their numbers, and extract the intersections of the lines as the supplementary points for the missing positions. Otherwise, connect the points horizontally and vertically directly according to their numbers and extract the intersections as the supplementary points.

Images