Multi-view Scene Abnormal Object Description

Abnormal object description of multi-view scene is to describe the abnormal object in multi-view scene, which is a challenging topic in the field of computer vision. Research on abnormal object description of multi-viwe scene has important academic value and broad application prospects. And it also has an important significance for the exploration of human activities, national security and public safety. There are some difficulties of abnormal object description of multi-view scene. For example, the accuracy of the traditional algorithm for moving target detection in complex scene accuracy is not high; the object always has more complex behavior; multi-view scene has a great influence on the abnormal target description; it is difficult to track and describe object in long time. In order to solve these problems, we carry out research on multi-view scene abnormal object description. The main work and innovation are listed as follows:1). A method of moving object detection based on CRF is proposed in allusion to the problem of low accuracy of the traditional moving object detection. The movement features and color features are extracted and then the feature vector is modeled by CRF and then abnormal object is described. The experimental results show that the error rate of this method is14.38%, less than traditional method such as frames subtraction with error rate81.34%, optical flow with error rate33.59%, Gaussian mixture model with error rate19.73%. The computation time of this time is less than optical flow and Gaussian mixture model and closed to frames subtraction.2). In order to use many types of features, a method of abnormal object description based on MCRF model is proposed. Several features subsets can be formed through more Features extraction. Then we made use of CRF model to each feature subset and got CRF units. Finally, we combined all the CRF units to produce MCRF model which was utilized to detect abnormal activity. The experimental results indicate that the accuracy rate of this method is better.3). A method of scene description based on latent topic model is proposed according to the problem of the low accurate of traditional methods. The several features are extracted and clustered to visual words by k-means algorithm. The visual words are divided into semantic topic distributions by pLSA model. And then, the semantic topic distributions are modeled by CRF to describe scene. The experimental results show that the accurate of this method is91.4%, better than the SVM and Bayesian.4). A method of abnormal object description based on blob and trajectory is proposed for the problem of traditional method with no consideration of the impact of the scene. The scene is described as blobs and trajectory of the object is extracted. Blobs and trajectory are combined to form feature vector. The feature vector is modeled by HMM to describe abnormal object. The experimental results show that this method can achieve a combination of semantic state of scene and object trajectory, and has a larger significance for describe abnormal object of certain scene.5). A method of abnormal object description of multi-view scene is proposed to solve the problem that the object of single visual scene can not be tracked and described in long time. The multi-view scenes are described as semantic states in accordance with the order of the camera location. Trajectory of the object is extracted. Semantic states of multi-view scenes and trajectory are combined to form feature vector. The feature vector is modeled by HMM to describe abnormal object. The experimental results show that this method can describe abnormal object of multi-view scene more accurately.