Human Behavior Recognition And Comprehension Based On Conditional Random Field

At present, video-based recognition and understanding of human behavior has becomeone of the concerned forefront directions in the computer field, due to recognition andunderstanding of human behavior have broad range of applications and potential economicvalue in intelligent monitoring, advanced human-computer interaction, virtual reality andcontent-based video retrieval and other fields, the researchers carried out extensive andin-depth of its research work, made some achievements, but there are many unresolved issues.This paper summarizes the domestic and the international research on recognition andunderstanding of human behavior, based on the problems of the current conditions randomfield models; we propose the concept of fraction labels, to establish the Fractal ConditionalRandom Field model. The main contents are as follows:This paper summarizes the research of behavior recognition, research methods andpractical applications at home and abroad, generalizes the problems of behavior recognition;In the behavior recognition process, there are image preprocessing, feature extraction,characterization methods, database selection to be summarized and in-depth analysis; Wehave an elaborate explain to the Condition Random Field model and this model’s infer,Meanwhile, we compare this model with other models, and analyzed its advantages anddisadvantages, and introduce the practical application.In order to solve the problem of a short video in target detection process, this paper,present background subtraction algorithm based on mixture Gaussian, the algorithm uses themixture Gaussian model to quickly get the background, then use background subtractionalgorithms to quickly get the target; this algorithm can effectively eliminate light effects,effective and direct extract target from a video, and establish a good foundation for thesubsequent feature extraction and behavior recognition.In the probability graph model, the traditional marking methods can not reflect theintegrity of the current class and direction issues, so we propose the concept of fraction labels,we can use fraction labels to test integrity and direction of a sequence, verify the accuracymark, and can give behavior direction about this sequence, and thus make effective sequencelabel test, solve the problem of marked deviation. As conditional random field can achieveglobal normalization, the marked deviation problem can be solved, and have good lifeapplications in line with the description and modeling of human behavior. Because amount ofcomputation of the fractional label is too large, we proposed hierarchical conditional random field model, the model is divided into two layers, the top firstly carry out and label the videosequence in an integer, get the corresponding video segment, give recognition results to thebottom, the bottom computer out fraction label of this video segment and test the output, theresult is a set of fraction labels, after that we analyze directionality and integrity of this set,verify this video label to meet the continuity of a behavior, and then correct the recognitionresults, thereby improve the recognition capability of human behavior.The model can accommodate the observed values overlap and long-distancecharacteristics dependence, and will be more subtle, precise in dealing with complex humanaction recognition. In this area, it has a strong advantage. To conversion for action over theframe, it has a very good correction and improves the robustness and recognition accuracy toa certain degree. The model compare with CRF, HCRF, LDCRF model to experiment, and theresults show that the algorithm has a good result in human behavior recognition.