Abstract:This paper presents a target detection algorithm based on PointRCNN. This method is aimed at vehicle targets. Aiming at the problem that the original PointRCNN is poor in vehicle detection at a distance, the method is optimized and the average accuracy of target detection is improved. In the first stage, the lidar point cloud is processed by pseudo-image structure and dimensionality reduction to 2D, and then processed by Point-Focus structure and restored to 3D point cloud. Then it will be sent into the backbone of PointNet++ for feature extraction, classification and regression. In the second stage, 3D frame is optimized and selected, and Point-CSPNet structure is introduced to further improve network learning ability and robustness. In this paper, the Focus and CSPNet structures of YOLO series algorithms are used for reference. The effective information in the original point cloud is fully extracted and the feature, gradient changes in the network operation are effectively integrated to improve the detection accuracy of the network. The average accuracy of the improved algorithm is improved from 81.10% to 81.74% in 3D scenes of KITTI dataset; and it is improved from 86.87% to 88.20% in BEV scenes of KITTI dataset, and the detection effect of vehicle targets in the far distance of visual effect has also been optimized to a certain extent, which has certain positive significance for further optimization and improvement of unmanned driving technology.

Abstract:Aiming at the problem of 3D irregular point cloud format and uneven density, a fusion of multi attention mechanism and pointrcnn network is proposed for 3D point cloud target detection. This experiment mainly improves the pointrcnn two-stage network respectively. Firstly, the channel attention and spatial attention mechanism are serially input to the distribution of each network layer in the first stage by adjusting and normalizing in batch to further quickly identify three-dimensional features; Secondly, the cross position attention mechanism is introduced into the second stage network to avoid the position deviation of the cross path, so as to further refine the three-dimensional target position for feature extraction. The experimental results on Kitti data set show that compared with pointrcnn detection network, the improved network improves the average mean accuracy (map) of car and pedestrian tests by 1.2% and 1.9% respectively. Therefore, the improved method not only solves the problems of irregular point cloud format and uneven density, but also ensures the detection accuracy.