Comparative study for feature detectors in human activity recognition

This paper quantifies existing techniques for feature detection in human action recognition. Four different feature detection approaches are investigated using Motion SIFT descriptor, a standard bag-of-features SVM classifier with x² kernel. Specifically we used two popular feature detectors; Motion SIFT (MOSIFT) and Motion FAST (MOFAST) with and without Statis interest points. The system was tested on commonly used datasets; KTH and Weizmann. Based on several experiments we conclude that using MOSIFT detector with Statis interest point results in the best classification accuracy on Weizmann dataset but MOFAST without Statis points achieve the best classification accuracy on KTH dataset.