Human-UAV interactive perception: Skeleton-based iterative perspective optimization algorithm for UAV patrol tracking of large-scale pedestrian abnormal behavior

This paper presents a system framework for UAV patrolling pedestrian abnormal behavior in public places based on human key point features and a skeleton based UAV large-scale pedestrian patrol viewpoint optimal with PID-iterative learning control algorithm is designed. The system framework provides early warning of abnormal pedestrian behavior through images obtained by UAV and human key point recognition by OpenPose. Aiming at the problem of recognizability of images acquired during UAV patrol, an algorithm inspired by iterative learning control for tracking large crowds with human torso and shoulder skeleton information is proposed and a PID-iterative learning control algorithm is designed to improve the control effect under the situation that pedestrian motion has an approximate repeatability. The algorithm is able to enhance the number of photographed pedestrians and increase the sustained observation time of their behaviors. Finally, the usability and effectiveness of the algorithm are verified by experiment results. The proposed method will contribute to the research of interactive perception technology in embodied intelligence.