混合体素与原始点云的三维目标检测方法

3D object detection method based on mixed voxels and original point clouds

针对目前基于点云的三维目标检测算法存在精度与速度不足的问题,提出一种具有较好检测能力的三维目标检测算法,分别在特征提取网络与点云表现形式等方面提出新的解决方法。在三维稀疏卷积网络后加入双通道注意力,通过空间注意力与通道注意力更有效地学习到多尺度语义特征并生成更高质量的初始建议;使用点云和体素的混合表现形式以类似残差网络的结构组成类残差点云融合模块,进而构成集合抽象模块,加强点云对建议细化的影响力,提高检测精度,同时改进点云采样策略提高算法检测速度;在网络训练中,使用多种数据增强方式,提高网络泛化能力。在KITTI数据集上进行实验,结果表明:提出的三维目标检测算法,汽车检测精度为84.94 mAP,自行车检测精度为67.41 mAP,在检测速度上相较原始网络提高37%,具有较好的检测精度与速度。

Aiming at the problem of insufficient accuracy and speed of the current three-dimensional object detection algorithm based on point clouds, this paper proposes a three-dimensional object detection algorithm with good detection abilities, and proposes new solutions in terms of feature extraction networks and point cloud expression, respectively. Firstly, the two-channel attention is added after the three-dimensional sparse convolutional network, and, through spatial attention and channel attention, the multi-scale semantic features are more effectively learned and higher quality initial proposals are generated;then, the hybrid representation of point clouds and voxels is used to form a fusion module of class residual point clouds with a structure similar to the residual network so as to constitute a set abstraction module, which strengthens the influence of the point clouds on the proposal refinement, improves the detection accuracy, and optimizes the point cloud sampling strategy to improve the algorithm detection speed;finally, in the network training, a variety of data enhancement methods are used to improve network generalization capabilities. Experiments on the KITTI dataset show that the three-dimensional object detection algorithm proposed in this paper has a detection accuracy of 84.94 mAP for cars and 67.41 mAP for bicycles, which is 37% higher than that of the original network in terms of detection speed, and has good detection accuracy and speed.