Human action recognition using PEM histogram

In this paper, a new action feature descriptor PEM (PCRM-EOH-MOH) is proposed for fast human action recognition. This descriptor is constructed based on three information channels: Pixel Change Ratio Map (PCRM), Edge Orientation Histogram (EOH) and Motion Orientation Histogram (MOH) features. A video sequence is first represented as a collection of PEM features. Then, video representations are constructed in PEM features and integrated with the linear Support Vector Machines (SVM) classification schemes for recognition. Finally, our method is tested on two challenging dataset: the KTH human action dataset and our own dataset including six types of actions. Experimental results demonstrate the effectiveness of our proposed method.