结合上下文信息和注意力的点云目标检测方法

Point Cloud Object Detection Method with Context Information and Attention Mechanism

针对基于体素的点云目标检测存在上下文信息利用不足和特征提取能力不足的问题,提出一种基于体素的3D目标检测模型(AA-PointPillars),通过引入三重注意力机制获得更鲁棒的体素特征,同时建立不同体素之间的依赖关系。在主干网络中,通过引入卷积注意力模块CBAM从通道维度和空间维度两方面增强对检测有帮助的特征,抑制背景噪声,得到更精确的检测结果。在KITTI数据集上的实验结果表明,与基准模型相比,算法的均值平均精度mAP在BEV(Bird′s Eye View)检测和3D检测上分别提升了1.15%和2.28%,证明了方法的有效性。

Aiming at the problems of insufficient context information utilization and insufficient feature extraction ability in voxel-based point cloud target detection,this paper proposes a voxel-based 3D target detection model(AA-PointPillars),which is more robust by introducing triple attention mechanism.voxel features of the rod,while establishing the dependencies between different voxels.In the backbone network,by introducing the CBAM module,the features that are helpful for detection are enhanced from the channel dimension and the space dimension,and the background noise is suppressed,which is beneficial for the detector to output more accurate detection results.The experimental results on the KITTI dataset show that,compared with the baseline model,the mean average accuracy mAP of car,pedestrian and cyclist was improved by 1.15%and 2.28%in Bird's Eye View(BEV)detection and 3D detection,respectively.