Subspace ensemble learning via totally-corrective boosting for gait recognition

Human identification at a distance has recently become a hot research topic in the fields of computer vision and pattern recognition. Gait recognition has most widely been studied to address this problem, because gait patterns can operate from a distance without subject cooperation. In this paper, a local patch-based subspace ensemble learning method for gait recognition is proposed. This method first incrementally selects the most efficient local patches from which local patch-based subspaces are learned to preserve the proximity relationships among instance triplets: instance i is closer to instance j than to instance k. Then these subspaces are combined to generate an accurate and robust classifier in the totally-corrective boosting framework. Unlike existing triplets based metric learning which is generally not able to deal with the high-dimensional data, the proposed method uses local information instead of global data to guarantee the feasibility of computation, and then the effective implementation method is proposed. To further improve recognition performance, Gabor wavelet-based feature pools are built for training the classifier. We compare the proposed method with the recently Appearing in Proceedings of the 30 th International Conference on Machine Learning, Georgia, Atlanta, USA, 2013. Copyright 2013 by the author(s)/owner(s). published gait recognition approaches on USF HumanID Database. Experimental results indicate that the proposed method achieves highly competitive performance against the state-of-the-art gait recognition approaches.