Human activity recognition based on mid-level representations in video surveillance applications

Abstract

Human Action Recognition methods have prospered during the last decade. They seek to automatically analyze ongoing activities in different camera views by using machinelearning algorithms in video sequences. Various human action recognition methods match local features and global features using action class labels in which abundant visual spatio-temporal information can hardly be generalized. To overcome this problem, we propose a novel notion of mid-level representations to construct a discriminative and informative semantic concept for human action recognition. This work introduces a mid-level representation based on the Optical Flow (OF) method, Hu and Zernike moment together. First we extract from each video, U h and Uv motion vectors by forming motions curvatures. Second, we determine the Hu moment and Zernike that serve as the feature vector of an action. Our method was tested and evaluated through a classification of the KTH and Weizmann datasets, with an Artificial Neural Network classifier (ANN). The results prove the accuracy of the suggested approach.