FFL-Track: Classroom Students Pose Tracking with First and Latest Frames Based on Fliter

Abstract

Aiming at the problems of low accuracy and time-consuming, combined with the characteristics of students' mutual obstruction and small activity space, a student pose tracking model based on filter with first and latest frames in classroom video is proposed, established a 14-point human pose representation module to represent the unobstructed joints’ information with fewer parameters, in order to reduce data redundancy and improve the model calculation speed and tracking accuracy, the human pose filtering module was constructed to remove the invalid pose that detects errors and detect duplicate in-frame redundant poses, and through the FL-Track module is used to match the target students with information of the student's first skeleton and the latest frame to comprehensively calculate the similarity of the students in different frames by using IOU and greedy algorithms. the experimental results on the classroom student pose tracking dataset created show that in the same experimental environment, compared with the SOTA, FFL-Track’s accuracy is increased by 5.94\%, and the frame rate is increased by 1.2fps, which greatly reduces the reasoning time and provides a great foundation for the downstream tasks in the classroom scene.