A spatio-temporal graph attention association model for simulated multitarget tracking
TL;DR
A simulation-based study of a spatio-temporal association model that combines an LSTM trajectory encoder, graph attention-based spatial context aggregation, a weighted binary association loss, a motion-prediction auxiliary head, distance-gated candidate generation, and global oneto-one online assignment finds the learned model remains competitive on some seeds but does not consistently outperform the kinematic baseline.
Abstract
Reliable online multi-target tracking requires associating noisy detections with existing tracks while preserving object identities through missed detections and clutter. This paper presents a simulation-based study of a spatio-temporal association model that combines an LSTM trajectory encoder, graph attention-based spatial context aggregation, a weighted binary association loss, a motion-prediction auxiliary head, distance-gated candidate generation, and global oneto-one online assignment. The evaluation uses five random seeds, an association score threshold of 0.3, and diagnostic analyses for gate/context threshold sensitivity, protocol exposure bias, motion-loss weighting, and component ablation under a unified online-history protocol. Across five seeds, the learned online model achieved MOTA = 0.7131 +/- 0.0461 and real-target recall = 0.8067 +/- 0.0527, while the constant-velocity Kalman filter with gated Hungarian assignment achieved MOTA = 0.7535 +/- 0.0552 and real-target recall = 0.8490 +/- 0.0629. The learned model therefore remains competitive on some seeds but does not consistently outperform the kinematic baseline. The discussion distinguishes supervised training pair accuracy from online tracking performance and identifies exposure bias, score calibration, class imbalance, and gate design as the main sources of the observed gap.