基于金字塔特征融合的二阶段三维点云车辆检测

针对三维点云鸟瞰图特征提取不充分导致车辆目标检测性能欠佳问题，本文提出一种基于金字塔特征融合的二阶段三维点云车辆目标检测算法。首先通过降维处理并利用体素占用编码原始三维点云，得到二维特征图输入；然后，利用上采样网络传递高层语义特征，下采样网络传递低层位置特征，构建一阶段金字塔网络结构提取车辆目标特征；最后，通过候选区域提取层得 到不同尺度的候选区域，利用兴趣区域池化层对齐各候选区域尺度，并采用全连接层融合多尺度特征，提取不同感受野下车辆目标特征；此外，在损失函数方面，补充正余弦角度损失并加权到总损失函数中，优化车辆目标航向角预测。基于KITTI公开数据集的实验分析表明，本文算法相较基准网络能够有效补充三维点云鸟瞰图特征提取，在不同难度的检测任务中平均检测精度提高 了5.07%~8.59%。

Pyramid-feature-fusion-based Two-stage Vehicle Detection via 3D Point Cloud

To improve the performance of vehicle target detection in three dimension (3D) point cloud bird eyes view (BEV), this paper proposes a two- stage 3D point cloud vehicle target detection framework based on the pyramid feature fusion. First, the original 3D point cloud is encoded by dimension reduction and voxel occupancy, which results in a two-dimension (2D) feature map. Then, the up-sampling network is used to transfer high-level semantic features, and the down-sampling network is used to transfer low-level location features. A one-stage pyramid network structure is constructed to extract vehicle target features. The candidate regions with different scales are obtained through the region proposal layer. The scale of each candidate region is aligned by the region of interest pooling layer, and the multi-scale features are fused by the full connection layer to extract the vehicle target features under different receptive fields. In addition, in terms of loss function, the sine and cosine angle loss is supplemented and weighted into the total loss function to optimize the prediction of vehicle target heading angle. The experimental analysis based on Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) open dataset shows that the proposed algorithm can effectively supplement the feature extraction of 3D point cloud aerial view compared with the benchmark network, and the average detection accuracy in difficult detection tasks is improved by 5.07% to 8.59%.