GNN-Guided Track Branch Formation For Multiple Hypothesis Tracking

Abstract

Multiple Hypothesis Tracking (MHT) has shown its powerful performance in multiple target tracking applications. To reduce the number of formed track hypotheses, MHT usually defines three correlation gates called extrapolation, initiation and correlation gate, and create track hypotheses based on the observation falling in the gate. However, in closely spaced multitarget scenarios, observations of new target may fall within the initiation gate of existing track for a long time, leading to the difficulty of track initialization. Similarly, the track may also break due to the gating rules. In this paper, we propose to utilize global nearest neighbor (GNN) association result to guide track branch formation in MHT and format track branches based on the GNN association result. Simulation experiments show that the proposed GNN-guided approach is able to solve the above problems and controls the computational complexity well.