Multi-camera People Tracking With Mixture of Realistic and Synthetic Knowledge

This paper presents a solution for Track 1 of the AI City Challenge 2023, which involves Multi-Camera People Tracking in indoor scenarios. The proposed framework comprises four modules: Vehicle detection, ReID feature extraction, single-camera multi-target tracking (SCMT), single-camera matching, and multi-camera matching. A significant contribution of our approach is the introduction of ID switch detection and ID switch splitting using the Gaussian mixture model, which efficiently addresses the problem of tracklets with ID switches. Furthermore, our system performs well in matching both synthetic and real data. The proposed R-matching algorithm performs exceptionally well in real scenarios despite being trained on synthetic data. Experimental results on the public test set of 2023 AI City Challenge Track 1 demonstrate the efficacy of the proposed approach, achieving an IDF1 of 94.17% and securing 2nd position on the leaderboard. Codes will be available at https://github.com/nguyenquivinhquang/Multi-camera-People-Tracking-With-Mixture-of-Realistic-and-Synthetic-Knowledge

1. Introduction

Multi-Target Multi-Camera (MTMC) tracking approaches pose significant challenges as they require solving multiple computer vision problems, including person detection, single-camera multi-target tracking, and person re-identification. These challenges involve addressing variations in camera resolution, distance, view angle, non-overlapping camera views, crowded areas, and changes in illumination. Furthermore, due to the limited availability of labeled datasets, multi people multi-camera tracking remains a challenge.