Objective Human action recognition is an interesting but challenging task for unconstrained videos with complex background, illumination variation and camera motion. In this paper, we present an improved dense trajectory-based algorithm to improve the accuracy of human action recognition. First, dense optical flow is utilized to track the scale invariant feature transform key-points at multiple spatial scales. The histogram of oriented gradient,histogram of optical flow, and motion boundary

more »

... ogram are used to depict the trajectory efficiently. Second, eliminating the influence of camera motion based on trajectory direction consistency. The purpose of this operation is to improve the robustness of trajectories. Finally, the Fisher vector model is utilized to compute one Fisher vector for each descriptor separately, and then the linear support vector machine is employed for classification.Experimental results on KTH and YouTube datasets demonstrate that the proposed algorithm can effectively recognize human actions.