MACHINE PERCEPTION AND ROBOTICS GROUP

機械知覚&ロボティクスグループ
中部大学

Object detection, object tracking and action recognition are the building blocks for understanding human behavior. Of the three, object tracking plays a vital role in focusing on the person thus enabling action recognition. In this paper, we propose a tracking framework called ”Online Real Boosting”, an improvement over the popular Online Boosting. Online Boosting is a learning algorithm that selects a discriminative classifiers based on the just arrived samples, in a tracking scenario occlusion or appearance changes results in errors that propagate resulting in drifts. The proposed method reduces drift utilizing Real Adaboost and a probability density function of the object and background. The proposed tracking algorithm was trained for tracking the human head, and the results were compared against an existing method, mean-shift and Online Boosting. As a result, proposed method achieved comparable or better performance, besides improvement on processing speed by reducing the number of weak classifiers by a half compared to Online Boosting. Of the many challenges in object tracking, appearance changes owing to the articulated nature of the object is the biggest. In Online Boosting based tracking, weak classifiers are selected from a pool of classifiers trained offline, and that are sensitive to appearance changes. In this paper, a new feature structure that is robust to appearance change, called the ”Soft Decision Feature” is also introduced. The Online Real Boosting and Soft Decision Features were applied to snippets of humans with complex appearance changes. Experimental results show that the proposed combination tracked scenes with variations to human pose successfully while the other methods either drifted or failed.