Enhanced multi-object tracking via appearance feature gating using UMAP
Ken Arioka, Yuichi Sawada · SICE Journal of Control Measurement and System Integration · 2025
This study introduces a novel approach to enhance multi-object tracking performance in computer vision by addressing the challenges of inaccurate object matching in data associations. The core idea is to leverage the Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction to analyze high-dimensional appearance features, facilitating comparisons in a 2D space. By evaluating the similarity between objects in existing tracks and newly detected candidates, our method effectively identifies and corrects incorrect data associations. Unlike other methods that require pre-trained appearance feature models, our approach supports online comparison without the need for pre-trained. The effectiveness of the proposed method is demonstrates through experiments using the MOT17 and MOT20 benchmark, which show improvements in several evaluation indicators. Additionally, the method proves robust in handling complex scenarios such as occlusions, ensuring reliable tracking outcomes.