A network based on affinity matrix and attention mechanism for MOT
Zhu Zhendong, Ming Zhao · 2024
The current multi-object tracking task faces challenges such as target scale variation, occlusion between objects, and diverse motion patterns. The algorithm consists of a feature extraction module, an affinity estimator module, and a hierarchical data association module. The feature extraction module includes a ResVGG feature extraction network and a multi-head attention (MHA) mechanism. The ResVGG feature extraction network learns features of the detected targets at different scales, while the MHA selects features from different scales as inputs to the multi-head attention mechanism, which integrates shallow, middle, and deep features to globally model the targets and capture long-term dependencies between them. The affinity estimator calculates an affinity matrix by estimating the affinity of aggregated target appearance information. The data association module performs the tracking task using the affinity matrix and a hierarchical data association module based on the strategy of grading the affinity levels. Experimental results demonstrate that the proposed algorithm achieves significant performance on the MOT15 dataset with a MOTA of 42.38 and MOTP of 72.80, as well as on the MOT17 dataset with a MOTA of 55.65 and MOTP of 79.98. These metrics indicate that the algorithm can effectively handle issues such as occlusion, scale variations and diverse motion patterns in multi-object tracking.