关键词: 无人驾驶汽车, 环境感知, 多模态融合, 机器视觉, 深度学习

Abstract: Environmental perception is an important part of self-driving cars and traffic participants (such as cars,pedestrians,cyclists) are the key detection targets.To solve the problem of low accuracy of lidar-only methods in identifying small targets (such as pedestrians and cyclists) due to the sparsity of point cloud,so combining the advantages of lidar and image in target recognition,a target detection algorithm based on multi-sensor fusion,PointPainting+,is proposed.This algorithm,building upon the framework of the PointPainting algorithm,enhances the semantic segmentation stage by incorporating a strip pooling module.This optimization enables the algorithm to achieve better recognition capabilities for long bar-shaped objects.The experimental results demonstrate that PointPainting+ algorithm,compared to the PointPillars baseline algorithm,exhibits an average accuracy improvement of 9.14% for cyclist detection and 9.71% for pedestrian detection.The detection speed can reach 43 frames per second,which meets the real-time requirement,and this algorithm effectively improves the problem of poor detection of long-distance and small targets such as pedestrians and cyclists due to the sparse point cloud.