Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (20): 312-319.DOI: 10.3778/j.issn.1002-8331.2306-0225

• Engineering and Applications • Previous Articles     Next Articles

Point Cloud Dynamic Object Recognition for SLAM

DAI Ling, SONG Zhenbo, LU Jianfeng   

  1. Nanjing University of Science and Technology, Nanjing 210014, China
  • Online:2024-10-15 Published:2024-10-15

用于SLAM的点云动态物体识别

代领,宋振波,陆建峰   

  1. 南京理工大学,南京 210014

Abstract: Detecting and segmenting dynamic objects in a scene is crucial for building a consistent map. A dynamic object detection algorithm based on continuous point clouds is proposed to address the problems that current point cloud dynamic object detection algorithms rely on a large amount of data containing dynamic attribute annotations and limit the scanning method of LiDAR. The input for this algorithm consists of the point cloud to be predicted, adjacent frame point clouds, and pose information obtained through simultaneous localization and mapping (SLAM). A point cloud scene flow estimation algorithm is employed by this algorithm to estimate motion at a per-point level, and scene flow results are integrated using techniques such as point cloud clustering and principal component analysis (PCA) to infer instance-level motion information for the determination of dynamic object attributes. Furthermore, point cloud semantic segmentation is utilized as a plug-in to boost dynamic object recognition accuracy by categorizing points into movable categories. The need for annotated data with dynamic attributes for training is eliminated by the proposed algorithm, and it does not impose any restrictions on sensor scanning methods or the number of generated point clouds. Comparative analysis with state-of-the-art methods demonstrates the algorithm’s ease of training, high accuracy in object classification, and robustness.

Key words: simultaneous localization and mapping (SLAM), deep learning, point clouds

摘要: 检测和分割场景中动态物体对于建立一致性地图至关重要。针对当前点云动态物体检测算法依赖大量含有动态属性标注的数据、限制激光雷达扫描方式等问题,提出了一种基于连续点云的动态物体检测算法。将待预测点云、相邻帧点云以及通过SLAM(simultaneous localization and mapping)得到的位姿信息作为输入,利用点云场景流估计算法逐点估计移动情况,结合点云聚类、主成分分析(principal component analysis,PCA)等技术,整合场景流结果以获取实例级移动信息以判断物体的动态属性,并将点云语义分割作为判别点是否属于可移动类别的插件以提升动态物体识别精度。所提算法不需要具有动态属性的标注数据进行训练,并且对传感器的扫描方式、生成的点云数没有任何限制;与现有最先进的方法进行对比,具有易于训练、判断准确、结果鲁棒等特性。

关键词: 即时定位与地图构建(SLAM), 深度学习, 点云