Research On Human Action Recognition Based On Dense Optical Flow Trajectories

Human action recognition is an important research topic in the field of computer vision and is getting more and more attention in recent years. Because it has a wide prospect in the production and research of human life, such as video surveillance, human-computer interaction, video search and motion analysis. In this thesis, human action recognition based on trajectories is achieved. Especially the methods for dense feature points detection, dense trajectories construction and behavior modeling were studied, and the main contents of this thesis are as follows:(1) In the view of the problem that trajectories are built on the sparse feature points, the number of trajectories is small. Considering the dense sampling achieving success in the image classification, an improved algorithm of detecting dense feature points is proposed. When the image pyramid on the basis of multi-resolution is built, dense sampling grid is utilized. Points are detected by comparing the eigenvalues of the entire image and the dense grid. The experimental results validate that the proposed method is able to detect dense feature points in a variety of environments of the different datasets than the other feature point detection algorithms.(2) Aiming at the problem of some points tracking error using the median filtering in a dense optical flow field, a method for establishing trajectories based on the principal orientations of feature points is proposed. Each point is tracked to the next frame if its principal orientation keeps stable. When the amount of feature points in a trajectory reaches the threshold, a new trajectory will be built. Once a trajectory is available, feature extraction would be conducted. A pipeline containing spatiotemporal volumes is divided in temporal spatial domain. Each divided pipeline is short and its features are calculated in the form of average values, so these features are rotation-invariant. Experimental results show that features obtained by the improved method have better separability.(3) As for the behavior modeling, VLAD which is a new model representation in action recognition is used. It is more useful than BOF for high-dimensional features.The above works form a complete algorithm for human action recognition based on the establishment of dense and reliable optical flow trajectories. It can effectively improve the recognition results in Weizmann and KTH database, which shows the effectiveness of the improved algorithm in this thesis.