Autonomous Learning for Tracking and Recognition

Abstract

We present an efficient approach for autonomous learning an object model from video or image sequences. The idea is to employ online boosting technique to adaptively learn an object representation from only as few as one labeled training sample. Our main contributions are: (1) A robust updating strategy of a discriminative classifier, which allows effective learning of an object model for tracking and recognition; (2) Learning and tracking are performed in a single procedure with possibility of reducing drifting and ability to recover tracking failure; and (3) a simple yet reliable framework for object recognition. Our main concern is to use the approach for the problem of hand and face tracking and gesture recognition. However, the proposed framework can be applied to other objects. Experiments on different data sets (publicity available) show the efficiency of our approach over very recent published approaches on different objects.