Shape-based human action recognition using multi-input topology of deep belief networks

This paper proposes a supervised approach for human action recognition by taking the power of simple features and advantages of multi-input topology of deep belief network (DBN) for accurate classification. This method is based on motion energy image (MEI) and difference energy array (DEA) in order to describe spatial and temporal domains, respectively. MEI can makes difference between actions which have much inter-class variations. About similar actions in inter-class variations, a simple time domain feature named DEA is used. After creating feature arrays, multi-input DBN is used for accurate action classification. We assume camera view-point and background clutters are fixed and this technique is robust on poor video quality, different scales in videos and human or action variations such as differences in manner, speed, body size and clothing. Also, simple implementation is advantage of the proposed method. Experimental results show the superiority of this approach over competing techniques on KTH dataset.