An effective approach for human actions recognition based on optical flow and edge features

Automatic human actions recognition is an interesting and challenging problem, and impacts applications in domain such as human-computer interaction, surveillance, human actions retrieval system and robotics. Deriving an effective actions representation from image sequences is important step for successful action recognition. In this paper, we empirically evaluate actions representation based on statistical global and local features combination, optical flow and edge features, for human action recognition. Firstly, we extract Histogram of Oriented Optical Flow and Spatial Pyramid Histogram of Edge. Secondly, we create the discriminative features by using PCA and LDA. Lastly, we use ANN for actions classification. Our approach is systematically examined on KTH and Weizmann datasets. Extensive experiments illustrate that optical flow and edge features are effective and efficient for actions recognition. We observe our experiments that optical flow and edge features perform efficiently and robustly over a useful of video sequences or camera with static backgrounds, and yield promising performance in video sequences captured in real-world applications such as surveillance and robotics. In addition, we extract action features in 2D using Lucas Kanade and Canny algorithms that have low computational cost. This is feasibility to apply into robotics.