Geometric Invariant Model Based Human Action Recognition

Most of the state-of-the-art methods for action recognition are very complex and variant to the geometric transformation like scaling, translation and rotation. Cuboid based method required all frames to extract the cuboid of action that's why cuboid based methods are expensive. Other methods use contour based approach for feature representation which is not robust to noise. So we require a very fast and robust feature descriptor which is invariant to geometric transformations. To deal with the above challenges our approach employs a geometric invariant model based human action recognition. It uses R-transform for feature representation. From each video we require a limited (approx. 10-15) number of frames and after detecting normalized foreground, we apply R-transform on Reason of Interest. The features of R-transform are: it is invariant to RST (rotation, scaling and translation), robust to noise and its complexity is NlogN where N=size of image i.e. N=n(star)n. When we are using PCA and LDA for dimension reduction and ANN (Artificial Neural Network) for classification the accuracy of our method falls in between 90 to 96% and with the PCA and Euclidian Distance based Classifier it falls in between 87 to 92%.