Human Action Recognition And Alarm Implementation Of Abnormal Behavior

Human action analysis based on video sequences has become a hot research in computer vision field, it has been widely used in intelligent video surveillance, virtual reality(VR),motion analysis, video retrieval and other fields. Intelligent monitoring system can collect a large amount of video information everyday, how to optimize video information, classify human behavior accurately, detect and alarm abnormal behavior timely, which are needed for further research in this field. However, most of the current action recognition methods have a poor universality, which has better recognition performance only for some certain subjects,human action recognition technology is still to be deeply studied in the future.After analyzing and summarizing human action research at home and abroad, on the basis of existing human action recognition technology, this paper establishes their own action video library and proposes two different recognition method to identify four daily behavior about falling, walking, sitting down and bending over. One is based on temporal and spatial shape feature, the other is a fusion recognition method based on global and local features. In addition, this paper designs a set of real-time monitoring system software in the mobile client,when an exception occurs, the system can receive the alarm message timely.First, we extract foreground by using Gaussian mixture background modeling method integrated with three inter-frame difference method. And then, in terms of feature extraction and behavior recognition, the first method is to extract the Hu moment feature which is the MEI of motion video sequence by secondary methods, and combine Naive Bayes Classifier to classify and recognize. The second method is to extract SIFT feature and gradient direction feature of MHI of video sequence, and then followed by Bag of Visual Word theory to integrate them effectively, finally by Support Vector Machine to recognize and classify. At last, on the Android client real-time monitoring and alarm system design, we adopt Socket network communication to achieve the remote monitoring and alarm.Experiments show that the proposed approach has a high recognition rate for humanaction, it also has a better robustness, and when abnormal behavior occurs, the alarm function can be achieved timely.