Locality constrained dictionary learning for human behaviour recognition

Abstract

ABSTRACTImprovement of the accuracy and robustness of human behaviour recognition is still a challenging problem. In this paper, a human behaviour recognition method based on locality constrained dictionary learning (LCDL) is proposed, in which two global features, namely average motion energy image (AMEI) and enhanced motion energy image (EMEI), are employed for the description of human behaviours. A discriminative structured dictionary is learned by the LCDL algorithm, whose sub-dictionary corresponds to each type of human behaviour. Moreover, the sparse representation errors of the testing samples are used for the recognition. The results of simulations and performance comparisons on typical datasets show that the proposed methods can reduce the storage space and calculation quantity through the normalization treatment of AMEIs and EMEIs. Since the locality constrained conditions can enforce the intra-class distance and improve the discriminative ability of the structured dictionary, the recognition performance is improved.