Locality Constrained Dictionary Learning for Human Behavior Recognition: Using AMEI and EMEI

Abstract

In this paper, a human behavior recognition method based on locality constrained dictionary learning (LCDL) is proposed, the average motion energy image (AMEI) and enhanced motion energy image (EMEI) are utilized to describe the human behavior features. Using the LCDL algorithm to train sub-dictionary for each category of human behavior, and then cascading all the sub-dictionaries together to form a structured dictionary, the dictionary is discriminative. The sparse representation errors of the testing samples are used for recognition. Compared with existing methods, the proposed methods can reduce the storage space and calculation quantity through the normalization treatment of AMEIs and EMEIs. And the locality constrained conditions can enfore the inner-class distance and improve the discriminative ability of the structured dictionary. The human beavior recognition experiments on Weizmann and DHA datasets have proven the validity of the proposed method.