Disparity-augmented trajectories for human activity recognition

Numerous methods for human activity recognition (HAR) have been proposed in the past two decades. Many of these are based on sparse representations, which describe whole video content by a set local features. Trajectories, as mid-level features, capable describing movements interest points two-dimensional (2D) space. However, 2D trajectories might be affected viewpoint changes, potentially decreasing their accuracies. In this paper, we first propose and compare different trajectory-based algorithms recognition. Then, new way augmenting with disparity information, without calculation 3D reconstruction. Our obtained HAR results shown 2.76% improvement when using disparity-augmented trajectories, compared to classical trajectory information only. Furthermore, also tested our method challenging Hollywood dataset, competitive results, at much faster speed.