Adaptive GLMB filter with IoU-based birth modeling for UAV visual multi-object tracking

Abstract

This paper proposes an Intersection-over-Union-based Adaptive Birth Generalized Labeled Multi-Bernoulli (IoU-AB-GLMB) filter for UAV-based multi-object tracking (MOT), specifically designed to address challenges posed by small objects with indistinct appearance features and time-varying object numbers. The proposed method introduces an IoU-based adaptive birth probability estimation model, where detected bounding boxes are clustered using IoU metrics to analyze spatial relationships, allowing the identification of unassociated or weakly associated measurements as birth targets. Additionally, we also enhance the Gibbs sampling truncation strategy by incorporating hypothesis weights and target count constraints, enabling adaptive truncation to improve computational efficiency while maintaining critical track hypotheses. Built on the GLMB framework, our proposed filter provides a unified probabilistic model that handles detection uncertainty, target survival, birth, and disappearance through Bayesian recursion, eliminating the need for manually defined rules. Furthermore, instead of committing to a single optimal association, the GLMB filter retains multiple association hypotheses at each iteration, allowing for a more robust treatment of uncertainty. Experimental results show that IoU-AB-GLMB achieves MOT accuracy 41.29% and 39.07% on VisDrone and UAVDT. Despite not relying on appearance cues, our method performs comparably to state-of-the-art appearance-based trackers StrongSORT (43.06% on VisDrone; 27.93% on UAVDT). These results underscore the effectiveness of our algorithm in UAV tracking scenarios.