Action Recognition Using Sparse Representations

The developments in digital networks and sensors has lead to the generation of plethora of data. The present 'big data' era seeks methods which can process data efficiently. Action recognition is the research area with numerous potential applications. The crux of this research is processing voluminous data. Motivated from the success of sparse representations in image classification and face recognition, we extended this theory to action recognition. Experiments conducted on Weizmann dataset resulted in improved accuracy. Our goal now is to exploit sparse reparesentations to recognize actions in large datasets like UCF-101, HMDB51. We rely on principles of Compressed Sensing for robust classification with lower dimension features and choose convex optimization algorithms to solve sparse representations.



Sushma Bomma
Dr. Neil M Robertson