Multiple object tracker with integrated embedding-based optimization and occlusion-aware variants
Gonzalo Carretero, Juan Pedro Llerena, Luis Usero, Miguel Angel Patricio · Information Sciences · 2026
Multiple Object Tracking (MOT) is a challenging task that involves detecting multiple objects within a video sequence and maintaining consistent identities for each instance over time. Currently, MASA is one of the most promising methods for MOT, given its zero-shot tracking ability. However, its performance has been limited due to a lack of robustness against occlusions and re-identifications. This work presents the deployment and systematic enhancement of a modern state-of-the-art multi-object tracker, along with a rigorous empirical evaluation of the proposed modifications. The results demonstrate that incorporating classical optimization and estimation techniques into modern embedding-based tracking pipelines yields measurable improvements in association accuracy and robustness under partial or full occlusions, without compromising computational efficiency. Moreover, a novel Occlusion-Aware (MASA-OA) tracking variant is devised that dynamically balances visual embedding similarity with geometric proximity (Intersection-over-Union), improving resilience when appearance cues degrade under partial and full occlusive conditions. The proposed MASA-OA tracker yields substantial gains in association quality, specifically increasing Association Recall (AssRe) to over 98% across all MOT15, MOT16, and MOT17 datasets, in addition to an 11.5% improvement on average in Association Accuracy (AssA) over the baseline. Overall, the presented analysis establishes a strong foundation for advancing association-centric tracking and provides a transparent, benchmarked basis for subsequent developments. The official code repository for this work can be found at https://github.com/gonzalocarreteroh/MASA-OA • Proposes MASA-OA, an occlusion-aware variant of a state-of-the-art tracker. • Replaces greedy matching with Hungarian assignment for optimal association. • Integrates Kalman filtering and embedding freezing under medium/high overlap. • Delivers consistent HOTA/association gains across MOTChallenge.