Practical action recognition with manifold regularized sparse representations

With the explosion of long term health conditions, monitoring human daily activities in home environment is one of the important issues in healthcare. Human action recognition in videos is one of the main topics in this context. Conventional representations are not very effective for encoding dense features extracted from videos. In this work, we propose a novel manifold regularized sparse representation (MRSR) method to encode dense features for human action recognition in assisted living. The new method can effectively incorporate a manifold regularization term to explore the geometric structure of the improved dense trajectories, which are very effective for learning action representations. By introducing a locality constraint, our method ensures each interest point is represented by its local closest words. Moreover, our method has an analytical solution and low computational complexity. Experimental results on different realistic databases show the effectiveness of the proposed algorithm for practical action recognition in assisted living.