Robust visual tracking through deep learning-based confidence evaluation
au Euntae Hong, au Juhan Bae, au Jongwoo Lim · 2015
In this paper, we propose an object tracking method through deep learning-based confidence evaluation, aiming at correctly updating an object template and on-line training a deep neural network. Our method updats both a deep neural network and a detector in Tracking-Learning-Detection(TLD) framework by robustly finding object regions highly similar to the target. We detect tracking failure points by measuring spatiotemporal similarity from Forward-Backward Error and output of the deep neural network. In addition, the proposed method adaptively updates the templates of tracker by finding the region with highest confidence of neural network within both tracking and detection results. Our experiment results demonstrate the effectiveness of the proposed method in severe environmental changes.