Research On Human Behavior Analysis In Videos

Human behavior analysis has become a research hotspot in the fields of computer vision andpattern recognition. It plays more and more important role in intelligent monitor, medicine and videogames. In this paper, several major parts of a human behavior analysis system and improved methodsaiming at existent problems are studied. This paper deals with the following tasks:Firstly, aiming at acquiring results of human motion detection, human body contour and skeleton,a method which is based on frame difference method is proposed. In order to improve the results ofhuman motion detection, maximum suppression is adopted. Simulation results demonstrate thatimproved approach can suppress interference efficiently.Secondly, aiming at the improving of human motion tracking, four measures, including refiningthe updating of target features, correcting the value of prediction based on particle filter, adapting theweights of wide feature and color histogram, calling the database of human motion detection, arepresented. The final results demonstrate the validity of the improved method.Subsequently, a method based on non-negative matrix factorization (NMF) and particle filter ispresented. Inspired by the non-negative character of non-negative matrix factorization, NMF andblock non-negative matrix factorization (BNMF) are introduced respectively to detect shadow areas.As for human motion tracking, the feature of the initialization stage based on NMF and a modifiedweight calculation formula of the particles searching stage based on NMF are adopted. Experimentalresults indicate that the tracking accuracy is improved and the speed of method is satisfied.Following, considering the recognition of human motion behavior, a method which is based onwhole wide features is proposed. Measures, including adopting three parameters (maximum wide,middle wide, average wide) in discriminant function, extracting key frame images which are obtainedby using system cluster method, using a frame image to describe a class image, combing adjacentframe images which are the same class, are utilized to optimize algorithm. Simulation resultsdemonstrate that these approaches are feasible.Finally, in this paper, action sequence is presented to judge abnormal behaviors. Experiments areconducted by analyzing action sequence of test sample with the classes of actions. Simulation resultsshow that our approach not only classify abnormal behaviors, but also find the abnormal action inbehaviors.