Classification-lock tracking approach applied on person following robot

The task of following a person in the real complex environment by camera still keeps at risk even the visual tracking technologies have been well studied in the last decade. Currently, most approaches only utilize single-shot initialization in the first frame and update their tracking models according to the result of the last frame. However, it leads to an uncorrected target selection once the inner appearance changes, i.e., a feature-rich object is moved out of the human. In this paper, we reveal a classification-lock tracking framework and apply our approach on a mobile platform. A pairwise cluster tracker is used to locate the person. A positive & negative classifier is utilized to verify the tracker's result and update tracking model. In addition, a pre-trained CPU optimized neural network is employed to lock the tracking result to only be human. In the experiment, we deploy the common challenges of visual tracking both on the static scene and a real-following task. Furthermore, our approach is compared with other state-of-art approaches on common datasets. Results prove the tracking quality of our approach in both the static and the dynamic scenes. Our approach achieves the best average score on the common dataset.