Application of ROI Feature Decomposition and Local Matching in Multi-Object Tracking

Abstract

For multi-object tracking (MOT), matching newly detected target features and trajectory features often involves template matching. Typically, feature embedding for multi-object tracking relies on either the mean feature within the Region of Interest (ROI) or the vector at the target's center. However, both of these feature embeddings are vulnerable to interference in scenarios involving occlusion. Additionally, one-to-many (O-T-M) style matching strategies are prone to causing the loss of trajectories and detections, as one target only represented by a single indicator. To overcome those limitations, we propose utilizing Principal Component Analysis (PCA) on the target ROI to decompose target features, expressing the identical feature of targets by collectively representing the principal component features of the ROI region, and utilizing these principal component features to form local features, the identical features are then trained through a classification task. Furthermore, in the template matching process, we introduce a local features matching algorithm for targets and trajectories. This algorithm adopts a many-to-many (M-T-M) matching style, aiming to associate every detailed part of the ROI feature. Experimental results demonstrate that the proposed local matching algorithm effectively handles complex and dynamic scenes. The approach achieves leading performance on the motchallenge benchmark, validating its effectiveness.