The Application Of Recurrence Plot Analysis Method In Mining The Relevance Of Image Data In Video

Today, the intelligence is developing rapidly, behavior recognition has received extensive attention, and has become one of the important study contents in the field of computer vision. It can be widely used in the field of human-computer interaction, video surveillance and intelligent robot. While the view irrelevant behavior recognition gets the study of more and more researchers. this paper studies human behavior of five views in video, using the IXMAS database as the sample database, and in this paper, the recurrence plot based on self-similarity matrix is used to mining the relevance of image data that in video, and applied in view-irrelevant behavior recognition and gets good recognition effect.Human behavior is a very complicated nonlinear process, and chaos theory is a common analysis method for nonlinear dynamical systems. People have explored applying the chaos theory to the study of human behavior recognition, its main idea is extracting the chaotic invariants which represent the human behavior in the reconstructed phase space, to form the feature vector that can be used to describe human behavior. This demonstrates the feasibility of using chaos theory to recognize the human behaviors.The raise of traditional recurrence plot analysis(RP) method provides a new way for people to analyze the problem from the view of nonlinear dynamics. It reconstructs the one-dimensional time series to the information-rich high-dimensional phase space, and then draws a two-dimensional recurrence plot, it not only reflects the chaotic invariant features of nonlinear system, but also allows people to observe the system state in the high-dimensional phase space on a two-dimensional image. This traditional method uses 2-D matrix which consists of black and white dots to characterize the system features. Its texture information is not rich, and it is complex to choose the phase space reconstruction parameters. In this paper, we use recurrence plot analysis method based on self-similarity matrix(SSM) to mine the relevance of human behaviors from different views, because the low-level feature descriptor we use is high-dimensional data, it contains rich information of system, so we use self-similarity matrix to build recurrence plot in the original phase space directly. In this progress, we don’t perform the binary processing, compared with the traditional recurrence plot, this method is focused on the use of richer texture information.In order to mine the relevance of image data in video of same behavior from different views, in this paper we use the video segments of same behavior under five views as the learning sample and use Dollar method to extract the spatial-temporal interesting points(STIPs), referring to SURF thought and combine the information; around of spatial-temporal interesting point, builds descriptor of the spatial-temporal interesting point in three-dimensional space. Then compute the self-similarity matrix of descriptors of five views respectively in a frame-based unit to get the recurrence plot of each view, they reflect the relevance of image data. In order to characterize the relevance further, we use the gradient direction distribution vector to build recurrence feature descriptor of each frame in the video, thus get the recurrence feature data set S. To mine the relevance of data in dataset S and then realize the view-irrelevant behavior recognition, so we cluster the recurrence feature descriptors in dataset S using NERF C-Mean method to produce M key words, the number of key words relies on the specific behavior. Then build the Gauss mixture model(GMM) for each key word. At last, compute the Statistical probability of descriptors in S distributing in M key word bags and form the descriptor vector of the behavior as the behavior detection template. The progress of behavior recognition is:build the recurrence descriptors of each video for test and distribute them to the corresponding key word bags, then count up the distribution vector and compute the similarity distance compared to the template. if the distance that has been computed is less than the threshold, the behavior of video for test is same to the template, otherwise it is different. In this paper we mine the relevance respectively of walk, sit down and kick from different views, and do the view-irrelevant behavior recognition experiment, and the recognition rates are above 80%.