A LiDAR target detection method in rainy and snowy weather based on Mahalanobis distance and Pointpillars

Abstract

The present invention relates to a method for LiDAR target detection in rainy and snowy weather based on Mahalanobis distance and Pointpillars, which is applicable to target detection in harsh environments such as rainy and snowy weather. This method first filters out rain and snow noise points by combining Mahalanobis distance and Pointpillars, and then uses the Pointpillars algorithm to output target detection results. Specifically, LiDAR point cloud data is acquired, projected onto a two-dimensional plane, and a KD tree is constructed. The distance of each point is sequentially compared with a set minimum radius. Points within the radius are searched using a KNN or RadiusNN algorithm based on the Mahalanobis distance, and outlier filtering is performed. Finally, the processed point cloud is voxel-encoded and used as input for the Pointpillar detection algorithm, which then outputs target detection results after processing. Compared to traditional LiDAR target detection algorithms, this algorithm can improve the accuracy of pedestrian and vehicle target detection in rainy and snowy weather, while also offering faster computational speed.

Claims

1. A method for detecting targets by lidar in rainy and snowy weather based on Mahalanobis distance and Pointpillars, characterized in that: The method comprises the following steps: Step 1: Get a frame of (x, y, z, r) feature point cloud data; Step 2: Project the point cloud onto the xoy plane and construct a KD tree based on the x-axis and y-axis data. Step 3, traverse each point in a frame of point cloud and determine whether the distance is less than the search radius; that is, determine Relationship with SR size; Step 4: KNN searches for the nearest k points or RadiusNN searches for points within a radius; Step 5: Calculate and compare the Mahalanobis distance between the traversed point and the k points. If it is greater than the set Mahalanobis distance D, it is determined to be a rain or snow noise point; or compare the number of point clouds within the search radius. If the number of point clouds within the search radius is less than k, then the point is determined to be a rain or snow noise point. Step 6: voxel-encode the point cloud data after removing rain and snow noise; Step 7: Send the voxel-encoded point cloud to the Pointpillars algorithm to output the detection results; Among them, step 3 determines The relationship with the size of SR is as follows: for each point, if d is less than the search radius SR, the KNN algorithm in step 4 is used to search for the nearest k points in the kd tree constructed, that is, the comparison results with each node are used to access the KD tree downward until the leaf node is reached, and it is determined whether a backtracking operation is required. Finally, the coordinate index of the nearest k point cloud is returned, and the Mahalanobis distance between the traversed point and the k points is calculated and compared in step 5; If it is larger than the search radius, the RadiusNN algorithm in step 4 is used to search all points within the radius, and the number of point clouds is returned and the number of point clouds within the comparison range in step 5 is executed.

2. The method for detecting targets by lidar in rainy and snowy weather based on Mahalanobis distance and Pointpillars according to claim 1, characterized in that: In step 2, a KD tree is constructed. Specifically, for a two-dimensional plane, one dimension x-axis or y-axis is selected for division in the construction of each layer of the KD tree, and the k-dimensional data space is divided into two parts, and it is made as balanced as possible, and then recursively constructed. The method of constructing the KD tree is to determine the size of the x-axis and y-axis data, and construct a binary tree according to the rule that the left subtree is smaller than the root node and the right subtree.

3. The method for detecting targets in rainy and snowy weather using lidar based on Mahalanobis distance and Pointpillars according to claim 2, characterized in that: In step 5, the Mahalanobis distance of the point is compared with the k points. Specifically, if d is less than the search radius SR, the Mahalanobis distance of the k points is calculated and compared. The specific steps of the Mahalanobis distance are as follows: assuming that the mean of the k points is expressed as: The covariance is: The covariance matrix is: The Mahalanobis distance calculation formula is:

4. The method for detecting targets in rainy and snowy weather using lidar based on Mahalanobis distance and Pointpillars according to claim 3, wherein: In step 6, the point cloud data is voxelized. Specifically, the input point cloud data is first divided into networks according to the X and Y axes to form voxels, and the point cloud data is enhanced to form a tensor of D = 9, which are (x, y, z, r, x c ,y c ,z c ,x p ,y p ), where x, y, z, and r are the input point cloud data information, x c ,y c ,z c The center coordinate of the grid for the point cloud, x p ,y p It is the relative position of the point and the geometric center. Assuming that each sample has P non-empty voxels and each Pillar grid has N point cloud data, the sample is represented by a (D, P, N) tensor. Then, the MaxPooling operation is performed on the dimension where the Pillar is located to obtain a feature map of (C, P) dimension, and then P is converted to (H, W) to form a (C, H, W) tensor.

5. The method for detecting targets by lidar in rainy and snowy weather based on Mahalanobis distance and Pointpillars according to claim 3, characterized in that: In step 7, the Pointpillars algorithm is fed into the model, and the total loss function is defined as: THE cls =-α a (1-p a ) γ logp a , Where L is the total loss function, L cls is the loss function for classification, L loc is the loss function of the position, L dir is the direction classification loss function; β loc , β cls , β dir They are the coefficients of the position loss function, the classification loss function, and the direction classification loss function respectively; the parameter p is the probability that the sample is positive; α a is the weight coefficient, α and γ are the hyperparameters of the system; (x, y, z, w, l, h, θ) are the position of the prediction box and the length, width, height and heading angle parameters; SmoothL1 is the loss function; Δb is the offset between the predicted value and the true value.

Images