Robust tracking via monocular active vision for an intelligent teaching system

Abstract

The research of this paper investigates a practical intelligent tracking teaching system, addressing the problem of teacher detection and tracking via monocular active vision in real time. The split lines and position-based visual servo rules are created to realize the robust and stable tracking, which is designed to keep the tracked teacher in the middle of image with a fixed size by automatically controlling a pan/tilt/zoom monocular camera in either rostrum region or other regions in the classroom. Face tracking in rostrum region is initiated by a face detector based on Adaboost followed by a novel long-term tracking algorithm named as informative random fern-tracking-learning-detection (IRF-TLD), which has advantages for its high accuracy and low memory requirement using real-valued feature and Gaussian random projection. Moreover, Gaussian mixture model can be automatically started to detect the teacher's movement when face tracking fails or stand-up students are detected. Experimental results on many benchmark sequences, which include various challenges for tracking, such as occlusion, illumination and pose variations, and scaling, have demonstrated the superior performance of the proposed IRF-TLD method when compared with several state-of-the-art tracking algorithms. Extensive experiments in a series of challenging real classroom scenarios also demonstrate the effectiveness of the complete system.