Pose recognition initialized pedestrian tracking for drones
Xuran Tang, Shuhan Zhou, Dongdong Li, Xing Zhang, Yuxin Sun · 2024
Pose recognition and object tracking are two classical research topics in computer vision. Recently, pedestrian tracking with consumable drones gets a wide range of applications in aerial photography. In the initialization stage, pedestrian movement in the image causes inaccuracy of the pedestrian bounding box for the tracking algorithm. In this paper, we propose a pose recognition initialized pedestrian tracker for drones. This tracker runs on drone equipped edge computing device. First, the skeleton keypoints are detected by the OpenPose convolutional neural network. Second, the pedestrian pose is recognized according to the spatial distribution of all skeleton keypoints and then triggers the pedestrian tracking algorithm. Last but not the least, the rectangular image region of the pedestrian is fed into the tracking algorithm which leads the drone and camera to follow the pedestrian. Experimental results demonstrate that the pose recognition initialized pedestrian tracker for drones proposed in the paper is more accurate and robust compared with the manually labeled bounding box initialized tracker.