Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (3): 209-217.DOI: 10.3778/j.issn.1002-8331.2109-0239

• Graphics and Image Processing • Previous Articles     Next Articles

3D Object Detection Algorithm Based on Raw Point Clouds

ZHANG Dongdong, GUO Jie, CHEN Yang   

  1. Army Engineering University of PLA, Nanjing 210007, China
  • Online:2023-02-01 Published:2023-02-01



  1. 陆军工程大学,南京 210007

Abstract: Aiming at the existing problems in 3D object detection, such as difficult data sampling, insufficient feature extraction, limited receptive field and low regression quality of candidate bounding box, based on 3DSSD 3D object detection algorithm, a single stage and anchor-free 3D object detection algorithm RPV-SSD(random point voxel single stage object detector) is proposed, which based on the raw point clouds. The algorithm is composed of five parts, namely, the random voxel sampling layer, the 3D sparse convolution layer, the feature aggregation layer, the candidate point generation layer, and the region proposal network layer. By aggregating the point-wise feature of the keypoints, the sparse convolution feature of voxel, and the BEV(bird eye view) feature, the category, 3D bounding box and orientation of the object can be predicted. Experiments on KITTI datasets show that the algorithm performs well on the whole. It can not only hit the target in the truth label and regress an accurate bounding box, but also infer the category and complete shape of the object from its incomplete point clouds, and improve the performance of object detection.

Key words: deep learning, raw point clouds, object detection, single stage, anchor free

摘要: 针对当前三维目标检测中存在的数据降采样难、特征提取不充分、感受野有限、候选包围盒回归质量不高等问题,基于3DSSD三维目标检测算法,提出了一种基于原始点云、单阶段、无锚框的三维目标检测算法RPV-SSD(random point voxel single stage object detector),该算法由随机体素采样层、3D稀疏卷积层、特征聚合层、候选点生成层、区域建议网络层共五个部分组成,主要通过聚合随机体素采样的关键点逐点特征、体素稀疏卷积特征、鸟瞰图特征,进而实现对物体类别、3D包围盒以及物体朝向的预测。在KITTI数据集上的实验表明,该算法整体表现良好,不仅能够命中真值标签中的目标并且回归较好的包围盒,还能够从物体的不完整点云推测出物体的类别及其完整形状,提高目标检测性能。

关键词: 深度学习, 原始点云, 目标检测, 单阶段, 无锚框