Research On The Approach Of Human Action Recognition Based On Mutil-features Fusion

Human action recognition technology in video can be widely used in many fields such as intelligent video surveillance, content based video retrieval and so on.It has great research value, is one of the hot research topic with great challenging in computer vision field. In this paper, based on the research of development status and action recognition method existing in domestic and abroad, we learn and research the two key problems of action recognition: action representation and action classification.Action representation is to extract features which describe the characteristics of the action in video, this paper firstly extract trajectories which is based on SURF(Speed Up Robust Features) points matching between consecutive frames. The method based on trajectory avoids the traditional motion segmentation step. And then two kinds of features based on trajectory are extracted, one describes the information of trajectory itself by gathering the statistics of size and direction of displacements between adjacent points in a trajectory; Another is the spatio-temporal context feature of a trajectory, extracting the appearance and motion features in space-time volumes which is centered by a trajectory, and use HOG(Gradient Orations Histogram)describes appearance information, two descriptors HOF(Histogram of Optical Flow) and MBH(Motion Boundary Histogram) describe motion information, thus extract the spatio-temporal context information of a trajectory.As to Classification of actions, this paper uses the bag of words model and support vector machine to complete the classification problem of action. First built a dictionary for every descriptor, and then use dictionaries built for quantification the video features to histograms, in this process, in order to solve the problem of lacking spatial and temporal information of bag of words model, we divide the video into several space-time blocks, use multichannel bag of words model about which we use descriptor and space-time block tag of each channel, Establish the histograms for each channel, and connect the histogram of all the channels of a video to establish a histogram for the video. This can not only increase the spatial and temporal characteristics, but also more flexible to fusion various features extracted above. At last, input the histograms of every video into SVM for training and testing, finishing the process of action classification.Finally, validate our action recognition algorithm on a plurality of classic video datasets. Experiments results show that, the correct rate of action recognition in this paper can not only improve the recognition accuracy on videos with simple background and simple action, but also can improve the action recognition accuracy on videos with complex background, and complex actions, proved that our action recognition algorithm is effective.