Abstract: In view of the problem that the 3D target detection Point cloud processing method has fewer target points under sampling and the loss of small target feature information, an improved Point-Voxel feature extraction method is proposed: First of all, based on the current advanced 3D target detection PV-RCNN model, a C-FPS (central-farthest point sampling) algorithm is proposed to solve the problem that the number of target points is small after point-based sampling. That is, the point cloud range is screened by images, and a normalized multiplied by the offset of the center point is added to X according to label Settings to optimize the point cloud distribution. Increase the number of down-sampling target points; Then, aiming at the balance between voxel-based voxel size division and feature extraction, a voxel image feature fusion method was proposed. The target image features were extracted by multi-channel convolutional neural network, and the multi-channel features were fused with the point cloud features extracted by Voxel-based to make up for the loss of feature information caused by the division size. Finally, validation is performed on KITTI data set. Experimental results show that compared with PV-RCNN model, the feature extraction method can effectively improve the detection ability of small targets in the current computer vision problems, and the average recognition accuracy of small target pedestrians and cyclists is better than that of PV-RCNN model, with the improvement of 1.62% and 1.81% respectively.