Improving Multi-Object Tracking performance via Graph-Augmented Detection under Computational Constraints

Real-time Multi-Object Tracking (MOT) methods have achieved remarkable progress. However, these methods still face significant challenges when deployed in computationally constrained environments, such as robots and other unmanned systems, where achieving both high detection accuracy and temporal stability remains difficult. A straightforward approach to improve this is to incorporate temporal information into the lightweight detector. Although many frameworks have been proposed, most of them are either restricted to offline processing or computationally expensive for practical application. In this paper, we introduce a novel and simple Graph-Augmented Detection (GAD) approach that leverages topological relationships to improve MOT performance under real-time and resource-constrained deployment. Additionally, we propose two graph augmentation strategies, enabling the model to be trained in a manner consistent with standard object detection, while maintaining stability and simplicity. Experimental results on the MOT17 and DanceTrack datasets demonstrate its effectiveness, faster training convergence and notable performance improvements. Moreover, our method incurs only 0.5% extra parameters and a modest 10% increase in inference time, offering a compelling trade-off between accuracy and efficiency compared with other information-incorporated methods.