Action Recognition Via Multi-features Fusion

Being one of the emerging disciplines, computer vision recently developed veryrapidly. Human behavior recognition as the key technology of video analysis and understanding is widely used in intelligent video surveillance system, human-computer interaction, video retrieval, medical diagnosis and care, etc.With fast access to the Internet and the wide spread of multimedia, video data because of its rich expressive power, has been becoming an integral part of people's daily lives.On the other hand, with the popularity of smart home and the promotion of smart cities, video surveillance systems are generating large amounts of video data per day.How to effectively identify the behavior of the video becomes a hot topic in the field of video processing.However, the current human behavior recognition technology has many difficulties, such as high computational complexity, low universality, still at a low level stage.After summarizing and analyzing relevant research at home and abroad, this paper studies the behavior description of different scenarios,and looks for a description of lower dimension and better robustness,which can be applied to different scenarios and get higher recognition rate.A new behavior recognition framework, based on multi-features fusion using Motion Orientation Histograms(MOH) feature, 2D-Sift feature and Histograms of Oriented Optical Flow(HOOF) feature,is proposed.On moving region detection,we use inter-frame difference method integrated with Gaussian mixture background modeling to extract foreground,and presented a method based on pixel change rate map(PCRM),which detect moving region accurately.On the behavioral description,we split the gradient space into positive one and negative one,and reconstruct the MOH feature, which makes the MOH feature has direction invariance.According to the characteristics of human motion, the motion area is divided into the head, torso, legs three parts,and the HOOF feature is constructed.The introduction of optical flow direction density information is to overcome the shortcomings of traditional optical flow histogram features expressive capability.Codebooksrepresent for each behavior, are establishedaccording to Bag-of-Word theory,for 2D-SIFT feature and HOOF feature sparse representation. The use of Bag-of-Word approach deals with the problem of different behaviors have various number of interest points and achieves local features fusion effectively.Experimental results show that this method either in the Weizmann database or in the KTH database has a high recognition rate and robustness,where Weizmann has a number of more action but relatively simple environment and KTH has a more complex environment.