形状补全引导的Transformer点云目标检测方法

Shape completion-guided Transformer point cloud object detection method

针对雷达传感器采集到的场景点云中存在大量远距离或位于遮挡视角的形状缺失的低质量目标,其几何信息不足难以被识别,影响检测精度的问题,本文提出一种基于形状补全引导的Transformer点云目标检测方法(shape completion-guided transformer point cloud object detection method,STDet),通过增强低质量目标形状特征来有效提升目标检测精度,利用Pointformer主干网络提取场景点云特征以生成初始候选框,基于特征分离预测的形状补全模块重构候选框中残缺目标的完整形状点云;构建Transformer几何特征增强模型,融合目标完整形状信息及空间位置信息至各目标点特征中,并感知各目标点不同邻域掩码范围内的局部结构信息与全局几何特征的注意力相关性,以获取关键几何信息增强的目标全局几何特征;基于该特征引导生成精细化的目标检测框。在KITTI数据集上的实验结果表明,该方法在存在大量形状残缺低质量目标的困难场景中检测精度较基准算法提升了4.96%,大量消融实验证明了该方法所构建的形状补全算法和Transformer几何特征增强模型的有效性。

Aiming at the problem that in the point cloud of scenes collected by the LIDAR sensor,there are lots of lowquality objects with missing shapes due to long distance or occlusion,whose geometric information are too insufficient to be recognized,so that the detection accuracy is affected.Hence,a shape completion-guided Transformer point cloud object detection method(STDet)is proposed to improve the object detection precision by enhancing shape features of the low-quality objects.The features of the point clouds are acquired by the Pointformer backbone network to generate the initial candidate box.Then,the shape completion module predicted based on feature separation is designed to reconstruct a complete shape of point clouds of the incomplete objects within the candidate box.A Transformer geometric feature enhancement module is established,which integrates the complete shape information and spatial location knowledge of the object into its point-wise feature to perceive the attention correlation between the local structure information and the global geometric features within different neighborhood masks,so as to acquire the global geometric feature with enhanced critical geometric knowledge of the objects.Finally,the refined object detection boxes are generated under the guidance of global geometric features.Experimental results on KITTI data set show that compared with the benchmark algorithm,the proposed method improves detection accuracy by 4.96% in scenes with abundant low-quality objects of incomplete shapes.Meanwhile,the effectiveness of the proposed shape completion algorithm and Transformer geometric feature encoding module is proved by extensive ablation experiments.