Real-time 3D LiDAR Multi-object tracking for simultaneous localization and mapping (SLAM) in dynamic environments

Abstract

In general, a Simultaneous Localization And Mapping (SLAM) technique assumes a static environment with stationary landmarks throughout a robot's navigation. The detection and removal of dynamic objects from a scene is the most intuitive way to mitigate the degradation of a SLAM system in a highly dynamic environment. To identify dynamic objects, this thesis proposes a Light Detection And Ranging (LiDAR) based multi-object tracking technique employing an interacting multiple model Gaussian mixture probability hypothesis density (IMM-GMPHD) filter. First, ground-level scan points are removed, and the remaining points are clustered and represented by bounding boxes. The centers of the bounding boxes are tracked using an IMM-GMPHD filter and the mode probability of each bounding box is calculated. Lastly, after pruning and merging trackers, bounding boxes that correspond to trackers with high dynamic probability are discarded. While its time complexity is comparable to that of the state-of-the-art, the proposed method shows more than 80\% improvement in clustering performance, which is evaluated by newly-defined metrics. In addition, the tracking and SLAM performances are assessed in simulated traffic scenes, \textit{Highway} and \textit{City}, to demonstrate the effectiveness of the proposed algorithm in varying controlled environments. Last of all, dynamic object classification using the proposed method demonstrates the highest precision at the cost of a very slight decrease in recall in an extremely challenging real-world environment.