A feature edge-based single-frame detection method for vehicle orientation of laser point cloud

Abstract

The application relates to a laser point cloud vehicle orientation single-frame detection method based on a characteristic edge, and steps are as follows: S1, point cloud pretreatment, ground point cloud and high obstacle area point cloud which does not affect normal driving are removed; S2, the point cloud is clustered and the type is judged, and the clustered point cloud of the type of vehicle is extracted; S3, a convex polygon is constructed by using a Gramham scanning method, adjacent edges with an included angle greater than a set degree are optimized, adjacent edges with an included angle less than the set degree are combined, and a simplified convex polygon is obtained; S4, the edge of each simplified convex polygon is translated to the inside of the polygon by a set distance, points in all point clouds are projected upwards, and whether it is a characteristic edge is judged; and S5, the direction of the characteristic edge is used to construct an outer rectangle of the vehicle point cloud, compared with a true value, and an error result of vehicle orientation single-frame detection is obtained. Compared with the prior art, the application has the advantages of more accurate detection, lower error and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned driving, and in particular to a single-frame detection method for vehicle orientation based on feature edges. BACKGROUND

[0002] Methods for estimating vehicle pose based on LiDAR point clouds have been widely studied. Qu et al. addressed this problem using a horizontal mounted one-line LiDAR scanner, which is widely equipped on almost all autonomous vehicles due to its advantage in the field of view. They decomposed the L-shape fitting into L-shape vertex searching and L-shape corner positioning, which made their method computationally efficient. However, the laser points from the two-dimensional LiDAR scanner are sparse, and in many cases, they do not cover the entire L-shape boundary, which limits their detection range. Shen et al. proposed an algorithm that divided a cluster into two disjoint sets using the least squares error, and then fitted them with two perpendicular lines. The efficiency of this method is very high, with an average computation time of 1.49 ms per sample. However, their method is aimed at point clouds with relatively few points, and it is highly dependent on corner points, and they did not consider the case where only a single edge of the vehicle is detected, which is very common. To reduce the complexity of obstacle localization computation, Oniga et al. proposed a RANSAC-based method that generated a model consisting of two perpendicular lines and was fitted to a set of boundary cells. The line with more points around it was determined as the direction of the vehicle. The computation speed of this method is 0.2 ms per sample. To reduce the absolute average error, Benjamin et al. proposed a vehicle pose estimation method based on point cloud convex polygons, which reduced the average error to 2.78° in 2331 frame samples. To accurately estimate the pose of the vehicle from incomplete and sparse point clouds, Bo et al. developed an algorithm that takes into account the road normal feature. In 10 typical cases, the average directional error was reduced to 1.64°. Zhang et al. characterized the L-shape fitting as an optimization problem. Three criteria were proposed to estimate the direction of the vehicle, which are suitable for various types of point clouds. As an extension of Zhang's method, Yang et al. simultaneously calculated the direction and position of the vehicle based on the distance of the boundary box edges. Many researchers have made efforts to detect corner points because they are one of the most important features of the L-shape. Kim et al. proposed a computationally convenient geometric method to extract vehicle corner points. Using this method, Zhao et al. proposed a RANSAC-based L-shape fitting method to estimate the direction of the vehicle in multiple frames of a time series. Also using multiple frames, Wang et al. combined prior information about the road to effectively estimate the direction of the vehicle on the road.

[0003] Most of the model-based methods mainly focus on two characteristics of clustering point clouds: 1) points around the key edge are denser; 2) the corner of the point cloud is clear.

[0004] However, the dense edge may be the round edge of the vehicle head, which may cause errors in vehicle orientation detection, and the corner of the point cloud is not always clear, so the corner needs to be constructed, which may introduce new errors. SUMMARY

[0005] The purpose of the present application is to overcome the defects of the prior art and provide a laser point cloud vehicle orientation single-frame detection method based on feature edges.

[0006] The purpose of the present application can be achieved by the following technical solutions:

[0007] A laser point cloud vehicle orientation single-frame detection method based on feature edges, the method comprises the following steps:

[0008] S1, pre-processing the point cloud obtained by the laser radar, and removing the ground point cloud and the high obstacle area point cloud which does not affect normal driving;

[0009] S2, clustering the pre-processed point cloud, and judging the type of the clustered point cloud, extracting the clustered point cloud of the vehicle type as the object of algorithm implementation;

[0010] S3, using the Gramham scanning method to construct a convex polygon for the object of algorithm implementation, optimizing the adjacent edges with an included angle greater than a set number of degrees, merging the adjacent edges with an included angle less than a set number of degrees, obtaining a new edge by connecting the two ends, and obtaining a simplified convex polygon;

[0011] S4, translating each edge of the simplified convex polygon to the inside of the polygon by a set distance, projecting the points in all point clouds upward for each translated edge, and judging whether it is a feature edge;

[0012] S5, using the direction of the feature edge to construct an outer rectangle of the vehicle point cloud, comparing with the true value, and obtaining the error result of the vehicle orientation single-frame detection.

[0013] Further, the high obstacle area point cloud which does not affect normal driving includes road sign point cloud and leaf point cloud.

[0014] Further, the clustering of the pre-processed point cloud uses the Euclidean clustering method.

[0015] Further, the type of the clustered point cloud is judged by height information, size information and shape information.

[0016] Further, the set number of degrees is 170 degrees.

[0017] Further, the translation of a certain distance is specifically 8cm.

[0018] Further, the judgment of whether it is a feature edge is specifically: if a point satisfies the following two conditions:

[0019] a. The point is outside the edge;

[0020] b. The projection point of the point is within the range of the edge segment;

[0021] The point is considered to be controlled by the edge, and the feature edge is the edge with the most control points.

[0022] Further, the parallel direction or the vertical direction of the feature edge is the direction of the vehicle.

[0023] An electronic device comprising a memory and a processor, the memory having stored thereon a computer program, the processor implementing the method described above when executing the program.

[0024] A computer-readable storage medium having stored thereon a computer program, the program being executed by a processor to implement the method described above.

[0025] Compared with the prior art, the present application has the following beneficial effects:

[0026] Firstly, the present application focuses on the clear boundary characteristics of point cloud, extracts the feature edge, discovers the clear edge feature of vehicle point cloud, and uses a suitable modeling method to depict the feature, so that the detection of the edge is more accurate, and the detection error of the vehicle direction is reduced by about 10% compared with the prior art.

[0027] Secondly, the object of the method implementation is optimized to a simplified convex polygon, the calculation efficiency is optimized, and the speed is improved by about 50% compared with the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The flowchart of the method of the present application;

[0029] Figure 2 The schematic diagram for judging whether the edge of the simplified convex polygon is a feature edge, wherein d represents the translation distance, AB represents the edge before moving, and A'B' represents the edge after moving. DETAILED DESCRIPTION

[0030] The present application will be described in detail below in conjunction with the drawings and specific embodiments. The present embodiment is implemented on the basis of the technical solution of the present application, and gives a detailed implementation manner and specific operation process, but the protection scope of the present application is not limited to the following embodiments.

[0031] Embodiment

[0032] As Figure 1 and 2 shown, a laser point cloud vehicle orientation single frame detection method based on feature edges, which combines geometry and search methods, provides a new idea for vehicle orientation single frame detection, and provides more accurate detection results for subsequent multi-frame tracking, the method comprising the following steps:

[0033] S1, pre-processing the point cloud obtained by the laser radar, and removing the ground point cloud, road sign point cloud, and high obstacle area point cloud such as leaf point cloud which does not affect normal driving;

[0034] S2, using the Euclidean clustering method to cluster the pre-processed point cloud, and judging the type of the clustered point cloud through height information, size information and shape information, extracting the clustered point cloud of the type of vehicle as the object of the algorithm implementation;

[0035] S3, using the Gramham scanning method to construct convex polygons for the object of the algorithm implementation, optimizing the adjacent edges with an included angle greater than a set degree of 170 degrees, merging the adjacent edges with an included angle less than a set degree of 170 degrees, connecting the two ends to obtain a new edge, to reduce the influence of vehicle round edges and clustered singular points on the final result, and obtaining simplified convex polygons. If the feature edges are directly extracted from the convex polygons, two problems will be faced, one is that there may not be a feature edge representing the vehicle orientation due to the influence of singular points and vehicle round edges, and the error is large; the second is that the large number of edges will lead to longer calculation time, affecting the normal operation of the entire framework, therefore the convex polygons need to be simplified.

[0036] S4, translating each edge of the simplified convex polygon to the inside of the polygon by a set distance (d=8cm), i.e. moving from the position of edge AB to the position of edge A'B', for each translated edge, projecting all points in the point cloud upward to determine whether it is a feature edge; the determination of whether it is a feature edge is as follows: if a point satisfies the following two conditions:

[0037] a, the point is outside the edge (outside means opposite to the translation direction);

[0038] b, the projection point of the point is within the range of the edge segment;

[0039] then it is considered that the point is controlled by the edge, and the feature edge is the edge with the most control points; the parallel direction or perpendicular direction of the feature edge is the orientation of the vehicle.

[0040] S5, using the direction of the feature edge to construct the circumscribed rectangle of the vehicle point cloud, and comparing with the true value, the error result of the vehicle orientation single-frame detection is obtained. Through the test of the KITTI dataset, the RANSAC (random sample consensus) fitting method and the method of the application are compared, and the error result is shown in Table 1.

[0041] Table 1: Test results on KITTI dataset

[0042] Method Mean absolute error (degrees) Standard deviation absolute error (degrees) Based on RANSAC fitting method 2.56 1.75 The method of the invention 2.22 4.35

[0043] The application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to realize the method as described above.

[0044] The application also provides a computer readable storage medium, which stores a computer program, and the program is executed by the processor to realize the method as described above.

[0045] The preferred embodiments of the application are described in detail above. It should be understood that those skilled in the art can make many modifications and variations without creative labor based on the concept of the application. Therefore, any technical solution obtained by logical analysis, reasoning or limited experiment based on the prior art according to the concept of the application shall be within the protection scope determined by the claims.

Claims

1. A single-frame vehicle orientation detection method based on feature edges using laser point clouds, characterized in that, The method includes the following steps: S1. Preprocess the point cloud obtained by the lidar to remove the ground point cloud and the point cloud in the high obstacle area that does not affect normal driving. S2. Cluster the preprocessed point cloud and determine the type of the clustered point cloud. Extract the clustered point cloud of type vehicle as the object of the algorithm. S3. Use the Gramham scan method to construct a convex polygon for the object to be implemented by the algorithm. Optimize the adjacent sides with included angles greater than a set degree, merge the adjacent sides with included angles less than a set degree, and connect the two ends to obtain new sides, thus obtaining a simplified convex polygon. S4. Translate each simplified convex polygon edge into the polygon by a set distance. For each translated edge, project all points in the point cloud upwards and determine whether it is a feature edge. S5. Use the direction of the feature edges to construct the bounding rectangle of the vehicle point cloud, compare it with the ground truth, and obtain the error result of vehicle orientation detection in a single frame.

2. The method for single-frame vehicle orientation detection based on feature edges using laser point clouds according to claim 1, characterized in that, The point cloud data of high-altitude obstacle areas that do not affect normal driving includes road sign point clouds and foliage point clouds.

3. The method for single-frame vehicle orientation detection based on feature edges using laser point clouds according to claim 1, characterized in that, The clustering of the preprocessed point cloud uses Euclidean clustering.

4. The method for single-frame vehicle orientation detection based on feature edges using laser point clouds according to claim 1, characterized in that, The type of the clustered point cloud is determined by height information, size information, and shape information.

5. The method for single-frame vehicle orientation detection based on feature edges using laser point clouds according to claim 1, characterized in that, The set degree is 170 degrees.

6. The method for single-frame vehicle orientation detection based on feature edges using laser point clouds according to claim 1, characterized in that, The specified translation distance is 8cm.

7. The method for single-frame vehicle orientation detection based on feature edges using laser point clouds according to claim 1, characterized in that, The determination of whether a point is a feature edge specifically refers to: if a point satisfies the following two conditions: a. The point is outside the edge; b. The projection of this point is within the range of the side segment; It is then considered that the point is controlled by the edge, and the characteristic edge is the edge with the most control points.

8. The method for single-frame vehicle orientation detection based on feature edges using laser point clouds according to claim 7, characterized in that, The parallel or perpendicular direction of the feature edge is the vehicle's orientation.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images