基于PointPillars的无人驾驶汽车三维目标检测优化算法

Optimization Algorithm for 3D Object Detection of Self-Driving Vehicles Based on PointPillars

环境感知是无人驾驶汽车的重要基础。对于无人驾驶汽车进行三维目标检测时出现的检测精度低、检测速度慢的问题,提出了一种基于激光雷达的PointPillars改进的目标检测方法PointPillars+。该算法在原PointPillars框架的主干网络中加入协调注意力(CA)机制,能让网络模型专注图像中有效特征信息,忽视无效信息,从而提高检测的精度。之后在KITTI数据集中进行了验证,试验结果表明所提出的优化算法与原PointPillars基线算法相比,平均精度提升了1.24%,且检测速度依旧满足实时性能要求,达到60.0 fps。

Environmental perception is an important basis for self-driving vehicles.PointPillars+,a new method of target detection based on laser radar is proposed,in order to solve the problems of low precision and slow speed in 3D target detection of self-driving ve-hicles.The algorithm adds the CA attention mechanism to the main network of the original PointPillars framework,which makes the network model focus on the effective feature information and ignore the invalid information,for improving the detection accuracy.The experimental results show that the average detection accuracy of the proposed algorithm is 1.24%higher than that of the original Point-Pillars baseline algorithm,and the detection speed still satisfies the real-time performance of 60.0 fps.