改进Point-Voxel特征提取的3D小目标检测

3D small object detection based on improved Point-Voxel feature extraction

针对3D目标检测点云处理方法存在下采样目标点少,小目标特征信息丢失的问题,提出改进的PointVoxel特征提取方法.首先,以当前先进3D目标检测PV-RCNN(Point-Voxel Feature Set Abstraction for 3D Object Detection)模型为基础,就point-based (基于纯点云)采样后目标点数量较少的问题,提出C-FPS (基于中心最远点采样)算法,即通过图像筛选点云范围,根据标签设置对X增加一个归一化乘以中心点的偏移量,优化点云分布,提高下采样目标点数量;然后,针对voxel-based (基于体素)需要划分体素大小与特征提取平衡的问题,提出体素图像特征融合方法,通过多通道卷积神经网络提取目标图像特征,将多通道特征与voxel-based提取的点云特征进行融合,弥补划分大小导致的特征信息丢失;最后,在KITTI数据集上进行验证.实验表明,与PV-RCNN模型相比,在当前困扰计算机视觉中的小目标检测上,该特征提取方法有效地提升了对小目标的检测能力,对于小目标行人和骑行者,其平均识别精度均优于PV-RCNN模型,提升幅度分别达到了1.62%,1.81%.

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 infonnation,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.