Deep Residual Matrix Factorization for Gait Recognition

Abstract

Gait recognition is one of the hottest topic for biometric applications. By employing gait energy image (GEI), both static and dynamic information of gait images can be represented for distinguishable feature, yet with a high dimension. Most of the previous dimension reduction approaches devote to redaucing the dimension from original space to low-dimensional space directly, losing the temporary deep feature. To this end, this work proposes a deep dimension reduction framework named Deep Residual 2DPCA (DR-2DPCA) for gait recognition. To accelerate the computation, 2DPCA is utilized as atom operator. We achieve the reconstructed GEI with 2DPCA to boost the residual, and repeat this procedure according the pre-defined number of layers. According to decomposition of multiple layers, the discriminant feature can be achieved. Experiments on the CASIA-B gait dataset illustrates that the proposed algorithm outperforms the existing ones using traditional matrix factorization techniques.