Improved approach for action recognition based on local and global features

This paper presents an improved spatio-temporal (XYT) approach for local interest point-based global action representation, considering the history of moving points in an action. The presented spatio-temporal representation demonstrate robust results and we compare the developed method with previous other method. This is a SURF-based method where we extract visual features to select candidate points based on the SURF detector. Afterwards, motion features are extracted by exploiting the local interest points and by employing optical flow. RANSAC is employed to reduce the unwanted outliers and improve the performance of the method. Based on an outdoor action dataset, we have found that the developed method demonstrate satisfactory recognition results.