A Vision-based Slow-Fast Target-Positioning Framework for Person-Following

Abstract

The person-following technique has a promising prospect in living and industrial applications, but it encounters several practical issues such as scene changes and pose variations, especially when deployed on a mobile robot with limited computing power. In order to address the challenges in the person-following scenario, this study formulates the visual target-positioning procedure and proposes a Slow-fast Target-Positioning Framework (STPF) to provide robust target positions in real time. Within this framework, the fast branch enables the real-time capability, while the slow branch corrects the cumulative error and improves the robustness. A dataset is collected and setup to evaluate the impact of STPF on the performance of long-term person-following. Extensive experiments demonstrate that STPF reduces 75% interruptions compared to the KCF tracker baseline, and is well adapted to the long-term person-following in real scenarios.