IMM-SLAMMOT: Tightly-Coupled SLAM and IMM-Based Multi-Object Tracking
Zhuoye Ying, Hao Li
Abstract
In the context of autonomous driving systems, SLAM and dynamic object tracking represent pivotal challenges. Autonomous driving scenarios frequently demand the simultaneous acquisition of ego-pose and comprehensive motion information from the surrounding environment to enhance decision-making and scene comprehension.Given the inherent interdependence between these two challenges, a viable approach is to integrate SLAM and object tracking into an interconnected system referred to as SLAMMOT. However, many conventional SLAMMOT solutions rely on a single motion model for object tracking, which may inadequately capture complicated dynamics of real-world motions. In practice, object motion patterns can change from time to time, not conforming neatly to a single model. To handle existing challenges, this paper proposes the IMM-SLAMMOT, a tightly-coupled LiDAR-based SLAMMOT system that utilizes instance semantic segmentation and IMM modelling for dynamic object tracking. Ego-pose and dynamic object states are jointly optimized in an innovative graph optimization framework intimately integrated with the IMM algorithm. Comparative analysis against our baseline, which employs a single motion model for object tracking, demonstrates that the IMM-SLAMMOT outperforms at motion-pattern-transition moments and consistently achieves competitive results in SLAM and multi-object tracking tasks throughout the entire trajectory.