Dynamic Object Detection and Instance Tracking Based on Spatiotemporal Sector Grids

The presence of dynamic objects poses a challenge for LiDAR perception systems to understand spatial structures, especially in degraded indoor scenarios. Recent work on 3-D object tracking focuses on developing an accurate system without giving attention to low-dynamic objects. In this article, a real-time algorithm that integrates ground tracking, dynamic object detection, and instance tracking is proposed, to robustly accomplish multiobject tracking tasks. To balance accuracy and real-time performance, an efficient point-to-point-based ground tracking solution is introduced to estimate navigable areas for dynamic objects. The intrinsic spatiotemporal information of the point cloud map is exploited to design a novel real-time dynamic point cloud detection algorithm. Dynamic sector grids and the double-thresholding method are invented to enhance the robustness and accuracy of dynamic detection. Besides, dynamic instances are extracted from dynamic point clouds and their 3-D trajectories are tracked with multiperspective monitoring scheme. Due to the absence of labeled dynamic datasets for solid-state LiDARs, the dynamic point cloud of our self-collected dataset has been manually annotated in every LiDAR scan and the dataset will be open soon. The proposed method has been thoroughly tested with public and private datasets, and numerous experiments have validated its superior performance both in dynamic object detection and tracking.