A novel method for detecting video objects

Most existing video object detection methods utilize supervised learning to train a generalized detector to achieve good detection performance on various test datasets. However, when facing complex real world scenes, a trained detector may fail to detect some objects or produce several false alarms. In this paper, we propose an unsupervised incremental learning scheme to deal with such an issue. We first utilized a multi-instance learning process to construct appropriate loss function of Real Adaboost, and then present an online sample collection and processing techniques to improve the performance of incremental learning. Experiments demonstrate the effectiveness of our approach.