Generalized Label Grouping for Scalable Trajectory Estimation

Abstract

Multi-Object Tracking (MOT) is concerned with estimating trajectories from sensor measurements. MOT using the Random Finite Set (RFS) framework has been gaining popularity due to its rigorous mathematical foundation and versatility in applications. Notably, large-scale trajectory estimation can be successfully achieved by the label-partitioned Generalized Labeled Multi-Bernoulli (GLMB) filter framework. In this work, we propose an efficient method for grouping object labels in scalable GLMB filtering. Specifically, the label grouping problem for parallel computation is generalized by considering the intersection of predicted measurements, i.e., uncertainty regions. The proposed approach provides a flexible criterion to construct label graphs, whereupon a large number of object labels can be rapidly determined whether to be grouped or not. We demonstrate the performance of our method via large-scale data sets.