Action recognition from videos using sparse trajectories

In this paper, a novel, low-complexity, method for action recognition from videos is presented. A 3D Sobel filter is applied to the video volume resulting into a binary image with nonzero pixels in areas of motion. The non-zero valued pixels are spatially clustered using k-means and the most dominant centers of video motion are extracted. The centers are then tracked forming sparse trajectories, whose properties are later used to create a new feature type, namely the Histogram of Oriented Trajectories (HOT), describing the video. Feature vectors are finally passed to an AdaBoost classifier for classification. The proposed method reports competitive results in KTH and MuHAVi datasets, while remaining low in complexity and thus being suitable to be used in surveillance systems requiring low processing power.