Deep Learning based Pedestrian Tracking Framework using Body Posture Information
Wonyong Song, Kwan-Woo Lee, Sang-Chul Park · Korean Journal of Computational Design and Engineering · 2020
This paper presents a deep learning based (human) pedestrian position tracking methodology from a video. In a manufacturing shop floor, the moving trajectories of pedestrians can be utilized for various applications, such as layout optimization of facilities, and material flow analysis. To improve the accuracy of the proposed human tracking method, we make use of the body posture information of workers. The body posture information is extracted by using the ‘Open- Pose’ library which represents the first real-time multi-person system to jointly detect human body, hand, facial, and foot key points on single images. The proposed methodology consists of four major steps; 1) Pedestrian recognition from a video, 2) Pedestrian tracking by using the posture information, 3) correction of potential tracking errors, and 4) finding the moving trajectories of pedestrians. The proposed methodology has been implemented and test with various examples.