Aiming at the problems of low sampling efficiency and biased sampling accuracy of traditional PV-RCNN on the point cloud,an improved 3D target detection method based on PV-RCNN is proposed.The key point sampling strategy has been changed so that the limited key points can be more clustered within the scope of the proposal region and more effective foreground features are encoded to be used in the later proposal refine-ment,effectively generating more representative key points.We replaced the voxel set abstraction and the set ab-straction in the ROI grid pooling module with the VectorPool aggregation module for local feature aggregation to encode sparse and irregular point cloud data more efficiently.The algorithm is validated on the KITTI dataset,and the results show that the pedestrian bird's eye view detection takes a more significant difficulty level im-provement of 10.46％,and an overall frame rate improvement of 33.74％.Our method has better detection perfor-mance.
关键词
3D目标检测/卷积神经网络/点云/SPC关键点采样/VectorPool聚合模块
Key words
3D target detection/convolutional neural network/point clouds/SPC key point sampling/VectorPool aggregation module
维普
万方数据