Trajectory-based human activity recognition from videos

Sparse representation is widely used by different human activity recognition methods. Although many sparse feature extraction algorithms have been proposed in the literature, most of them focused on low-level features. This paper proposes a new method using trajectories, as mid-level features, for human activity recognition. Even though the use of trajectories is not new in this field, their potential is yet to be fully attained. In this paper, inspired by previous works, we have proposed new trajectory extraction methods, which are very flexible. Then we have emphasized the difference between trajectories and traditional descriptors, and have shown the advantages of using trajectories for human activity recognition. The pros and cons of trajectories are demonstrated through proposed trajectory-based methods. We have used a simple shape descriptor and the standard bag of word algorithm for human activity classification. The results of these different algorithms have been compared. We have also compared our results with other popular existing methods based on low level extracted features. In particular, we have shown that using proposed sparse trajectories can produce similar or better results than using dense trajectories. Furthermore, the computational time has been reduced as we are dealing with fewer data.