Local many-to-many matching via ROI feature decomposition for multi-object tracking

Jinlong Yang, Yandeng Ban, Jianjun Liu

Abstract: For multi-object tracking (MOT), data association often involves the process of matching appearance features. Typically, feature embedding for data association relies on either the mean feature within the Region of Interest (ROI) or the vector at the ROI’s center. However, these feature embeddings are vulnerable to interference in scenarios involving occlusion. To address this limitation, we propose decomposing target ROI features into a fixed number of vectors and subsequently applying Principal Component Analysis to select the top-k contribution vectors, forming local features. These local features serve two purposes. Firstly, utilizing these features to reconstruct an identical feature enables their training through a classification task. Secondly, these local features are employed in a Many-to-Many matching style that associates every detailed part of the ROI feature, thereby reducing mismatching. Experimental results demonstrate that the proposed local matching algorithm effectively handles complex and dynamic scenes. The approach achieves leading performance on the MOT Challenge public benchmark: MOT17 and MOT20.