Research On Human Behavior Recognition Method Based On Local Spatial Temporal Interest Points

Human behavior recognition based on video has always been concerned as a hot research topic of computer vision. It has a wide range of applications and utility value in all walks of life such as intelligent video surveillance, human-computer interaction, video retrieval and so on. The human behavior recognition method based on local Spatial-Temporal interest points has become the most popular method at present, since it has good robustness to various interference. This kind of method describes behavior by detecting interest points, whose pixel values in Spatial-Temporal neighborhood have a significant change, and extracting the underlying characteristics. Using this method, it is not necessary for background subtraction and object tracking. We do researches on this method, and the main contents and results of this paper are as follows:Propose a feature of HOIRM(Histogram of Oriented Interest Region Motion) based on ROI(Region of Interest Points). First of all, the space-time interest points area ROI according to the space position distribution of interest points in every frame is extracted, then the HOIRM can be determined according to the movement direction between the previous frame and the next frame. The HOIRM feature can be regarded as a kind of middle-level feature between global features and local features, it has the advantages of local features and avoids the difficult steps needed for the global features extraction as well. The local features are extracted by joining 3D HOG and 3D HOF feature description. By using the cumulative histogram, the local features and HOIRM feature can be effectively fused.Propose a BOW(Bag of Words) model based on AP clustering and apply it to human behavior recognition. The clustering effect of AP clustering algorithm itself is better than the K-Means clustering to form the BOW model on the recognition rate, when after multiple features fused. In the case of large number of features, AP clustering costs less time to obtain a visual dictionary. On the other hand, the BOW model based on AP clustering can obtain the optimal visual dictionary capacity by passing messages among feature vectors. In other words, it is not necessary to specify classification numbers and the initial clustering centers in advance. Thus, it is not necessary to do many experiments to get the proper visual dictionary capacity.