Point cloud 3D target detection method based on symmetric point generation

Abstract

The invention relates to a point cloud 3D target detection method based on symmetric point generation. The method comprises the following steps: firstly, filtering an original point cloud, voxelizinga detection space, generating initial voxel features, inputting the initial voxel features into a symmetric point generation module, obtaining high-level semantic features through a coding and decoding structure of the symmetric point generation module, and performing foreground point segmentation and symmetric point prediction through a classification head and a regression head; forming an enhanced point cloud by using the predicted symmetric point set corresponding to the foreground point and the non-empty voxel center point set as the input of the regional proposal network, further extracting the top view features through the backbone network of the regional proposal network, using the top view features as the input of the detection head, and finally outputting the 3D frame of the to-be-detected object by the detection head. According to the method, the symmetry of the detection object is utilized to generate the symmetry points, the problem of structural loss of the object in the point cloud is fundamentally relieved, the regression effect can be improved, the detection precision is improved, and meanwhile, replacement of RPN with other voxel-based detection methods is supported, so the original detector with a poor detection effect can generate a competitive detection result.

Description

technical field [0001] The invention relates to the technical fields of automatic driving and 3D target detection, in particular to a 3D target detection algorithm based on symmetrical point generation. Background technique [0002] 3D object detection has attracted more and more attention from industry and academia due to its wide application in autonomous driving, robotics and other fields. LiDAR sensors are widely used in autonomous vehicles and robots to capture 3D scene information in the form of point clouds and provide important information for the perception and understanding of 3D scenes. Since the point cloud can preserve the original size of the object, there is no problem of the object being too low resolution in the image, and the lidar can work normally even at night. Therefore, object detection in the point cloud scene has become a hotspot of 3D object detection. Currently, 3D object detection methods are mainly divided into two categories, one is image-base...

AI Technical Summary

Problems solved by technology

This method will rely heavily on the semantic information of the image, resulting in the inability to detect occluded or severely truncated objects

Images