Solving Tracking Challenges: Lightweight Algorithm for Single Pedestrian Tracking on Nano-UAVs

With its pocket size and flexible flight characteristics, the nano-UAV can achieve more stealthy flight. If the full on-board deployment of the single pedestrian tracking algorithm can be realized on the nano-UAV, It will completely unmanned tasks such as personnel monitoring and security inspections, and has a broad application prospects. However, it is very challenging to design single pedestrian tracking algorithms for nano-UAV platforms. On the one hand, nano-UAVs are not able to deploy the algorithms fully on-board due to their strictly limited computing and storage resources; on the other hand, there are various tracking challenges (multi-scale changing target tracking, small-scale target tracking) for single pedestrian tracking tasks. In this paper, we design a lightweight single pedestrian tracking algorithm that ensures its resource footprint meets the limitations of nano-UAVs. The algorithm proposed in this paper consists of a multi-task learning front-end and a fused ReID tracking back-end, and we design a complete algorithm deployment process, including dataset construction, model training quantification and model deployment, which can effectively solve the challenges of current tracking tasks. A high efficiency (computational efficiency reaches 4.8 MACs/Cycle, and energy efficiency reaches 4.7 mJ/frame) has been achieved on the nano-UAV, achieving an accuracy score of 0.988 and a success score of 0.578 with good robustness in flight. The full-board deployment of nano-UAVs can achieve real-time tracking efficiency (28fps), which has application value and greatly promotes the research work of nano-UAVs.