Human Activity Recognition using Optical Flow based Feature Set

An optical flow based approach for recognizing human actions and human-human interactions in video sequences has been addressed in this paper. We propose a local descriptor built by optical flow vectors along the edges of the action performer(s). By using the proposed feature descriptor with multi-class SVM classifier, recognition rates as high as 95.69% and 94.62% have been achieved for Weizmann action dataset and KTH action dataset respectively. The recognition rate achieved is 92.7% for UT interaction Set_1, 90.21% for UT interaction Set_2. The results demonstrate that the method is simple and efficient.