Research On Abnormal Activity Detection Based On Conditional Random Fields

Abnomal behavior detection is to detect abnormal behaviors take place in the scene, which is an important and challenging research topic in the field of computer vision. Study on the abnormal behavior detection method has the the broad application prospect and important academic value, playing an importanton role on national security, public security, family nursing.With the rapid development on image processing technology and pattern recognition technology recent years, abnormal behavior detection technology has made great progress. But there are still some problems about the complex behavior features, the low detection rate, and the abnormal behavior is difficult to define. In view of the above problems,we bagan our study on abnormal behavior detection based on conditional random field model. In this paper, the main research contents and innovative work can be summarized as follows:1) According the problem that the exist features extracted from the behavior sequence are difficult to describe human behavior fine, this paper proposed a feature extraction method which is based on3D skeleton model and body-based coordinate system. Considering the3D skeleton has the advantages of less data and retain the key information, the human body model is divided into5parts, then the body behavior is divided into global behavior and the behavior of limbs. To decribe the human behavior fine, we establish the body-based coordinate system, and extract mutilple features based on it. Also we have defined a proper standard human pose, with the standard human pose, we use the pose transformation parameters to describe the current human pose, so that the dimension of the feature vector is reduced.2) According to the HMM’s strict conditional independence assumption and the lack of context information modeling, we proposed a behavior recognition method based on conditional random fields (CRF) model. We compared the first-order CRF model with the HMM, experimental results show that the proposed method can get use of context information, to get a higher recognition rate.3) This paper proposes a behavior recognition method based on Mutiple Conditional Random Fields (MCRF) model. The advantage of MCRF model is the ability of combining more features and utilizing adaptive contextual information. First, human activity is divided into global activity, arm activity, and leg activity. Several features subsets can be formed through more features extraction. Then CRF models is used on each feature subset to generate CRF units. Finally, all the CRF units were combined to produce MCRF model which was utilized to recognize human activity. The experimental results indicate that the detection accuracy rate of this method is better.4) With the analysis of the hidden context informance among the human body parts, we propose a Star Conditional Random Field model (Star-CRF). We extend the structure of the first-order CRF to a star structure to jointly model the five human body parts, and introduced the features in the double position potential fuction. At last, we use the Star-CRF model to jointly recognize the human behavior. The experimental results show a high recognition rate and indicate that it can resist some self-occlusion effect.5) According to the problem that the abnormal behavior is hard to define, we proposed a idea of abnormal behavior detection based on the body behavior and the facial expression. With the idea, this paper proposed3methods to detect abnormal behavior. The firse method is to jointly model the body behavior and the facial expression with the CRF model, and to jointly detect the abnormal behavior. The second method is to take the body behavior feature and the facial expression feature as multi-feature, and use MCRF model to model them. The third method is to recognize body behavior and facial expression seperately, and use the two classification result to detect abnormal behavior.