Action recognition based on object tracking and dense trajectories

For recognizing human actions in video sequences, it is necessary to extract sufficient information which can represent motion features. In recent years, dense trajectories based action recognition algorithms attract more attention for containing rich spatio-temporal information. However, these algorithms are always faced with cluttered background. To solve this problem, we involve object tracking process in dense trajectories, with detecting object location through online weighted multiple instance learning object tracking and calculating dense trajectories in human bounding box, which suppresses redundant background information effectively. Meanwhile, human actions are classified by the bag-of-words model and SVM approach, which are robust to slight drift problem in tracking process. Involving object tracking improves efficiency as well as accuracy of dense trajectories. Our algorithm achieves superior results on the KTH and UCF YouTube datasets compared to the state-of-the-art methods, especially outstanding 89.0% accuracy on cluttered background dataset UCF YouTube.