Addressing the Occlusion Problem in Multi-Camera People Tracking with Human Pose Estimation

Jeongho Kim, Wooksu Shin, Hancheol Park, Jongwon Baek

Abstract: Multi-camera people tracking (MCPT) is a challenging task that is crucial for developing intelligent surveillance applications. In this work, we propose an MCPT system for Challenge Track 1 in the 2023 AI City Challenge. Specifically, we address the issue of occlusion, which causes significant changes in a person’s appearance and makes it difficult to estimate their exact location on a global map of a given area. In this paper, we present several solutions that utilize human pose estimation for overcoming this challenge. Our experimental results demonstrate that using human pose estimation significantly improves the performance of our system. Furthermore, we achieved promising results on the official evaluation set, with an IDF1 score of 86.76%. Our code is publicly available at https://github.com/nota-github/AIC2023_Track1_Nota.