Fast action detection with web camera

Although a great success has been achieved on action detection tasks by using "bag of features" architecture as video representations, action detection with web camera still remains a challenge. Most of these algorithms can extract features either sparsely at interest points or densely on regular grids, usually, sampling densely can get better results than sampling sparsely using the local descriptors. Here, we optimize the inputs information based on moving detection in the human visual system. Firstly, we employ change detection mask (CDM) algorithms to find informative regions which are estimated from the region getting by moving detection. Then, the feature descriptors, namely histogram of oriented gradients (HOG), are extracted corresponding to these mask regions. Finally, the Support Vector Machine (SVM) is used to recognize characters with pooling and encoding strategy. Experimental results show that the proposed method has a better performance even in the real world action detection using web camera and at the same time accelerates the speed of detection compared to the conventional method.