Robust Feature Extraction for Shift and Direction Invariant Action Recognition

We propose a novel feature based on optical flow for action recognition. The feature is quite simple and has much lower computational load than the existing features for action recognition algorithms. It has invariance to scale, different time duration and direction of an action. Since raw optical flow is noisy on the background, several methods for noise reduction are presented. Firstly, we bundle up the fixed number of frames as a block and take the median value of optical flow median flow. Secondly, we take normalization of histogram depending on the total magnitude. Lastly, we do low-pass filtering in frequency domain. Converting the time domain to frequency domain based on Fourier transform makes the feature invariant to shifted time duration of action. In constructing the histogram of optical flow, we align the direction of an action so that we can get direction invariant action representation. Experiments on benchmark action dataset KTH and our own dataset for smart class show that the proposed method gives a good performance comparable to the state-of-the-art approaches and has applicability to actual environments with smart class dataset.