A Study On Human Action Recognition In Video

The technology of human action recognition has been more important in the field of computer vision and widely used in the field of video surveillance, video and image retrieval, human-computer interaction currently. This thesis mainly studies the method of human action recognition. This recognition method uses a kind of valuable low-level features called Spatio-Temporal Interest Points(STIP) to represent human actions, creates action dictionary based on bag-of-visual words model(Bo VW) and learns action patterns by some action features. The goal in this thesis is to improve the recognition rate under complex scenes. The main works in the thesis are listed as follows:(1)The key technologies of human behavior recognition are discussed deeply. The methods of STIP detection was researched and analyzed, including 3D Harris and Dollar. The descriptors expressing STIPs information are studied, which are HOG, HOF and 3D HOG. Bag of visual words(BOVW) model is used as the representation of actions. Finally, human action is recognized by SVM classifier. The results show that these methods in KTH and UCF Sports database behave well.(2)For complex background will have the problem that some STIPs are not expected, this thesis proposes the method to increase the quality of STIPs. These STIPs is no helpful. It can suppress the negative effects caused by the complex background to extract the human body motion area where produce the excepted STIPs. This method will increase the quality of STIPs. It is the goal to ensure that the STIPs must be detected in the motion region. The results show that it can remove the bad STIPs and can improve recognition performance is also effective.(3)Focus on increasing the information of the descriptors, this thesis proposes two multi-feature fusion descriptors to describe STIPs and their neighborhoods. They are spatio-temporal fusion descriptor and trajectories-based fusion descriptor. The spatio-temporal fusion descriptor represents the structure information and motion information. The trajectories-based fusion descriptor represents more relative motion information than spatio-temporal fusion descriptor. The experiments show that the new descriptors do better on the behavior recognition and the trajectories-based fusion descriptor performs better.