The Research And Implementation Of Segmenting Groups And Detecting Anomaly In Crowded Scenes

In this paper, crowd segmentation and the abnormal behavior detection in dense scenes have been researched deeply. Group segmentation algorithm in crowded scenes is still in an immature stage. Currently, for surveillance video images of different vision field, different perspectives and different scenes, there is still not a standard unified method for group segmentation. Most of the algorithms are based on the macroscopic whole crowd or the microscopic individuals. Our segmentation method proposed in this paper is based on mesoscopic sub groups. This method can effectively improve the partition results under dense scenes by reducing failures caused by overlapping populations using microscopic methods and neglect of group behavior details using macroscopic methods. On the basis of this segmentation algorithm, this paper carries out the abnormal behavior detection based on the convolution neural network, and proposes a model based on group motion descriptors, which describe the intrinsic characteristics about crowd motion. This model obtains good results in experiments.Pedestrians in crowded scenes can be seen as collection of groups which moves coherently. In order to segment the crowd into groups that we can analyze the interactions between them and detect anomaly, we propose a novel method using group spatiotemporal relationships. Firstly, we acquire the spatiotemporal information of crowds by modeling the background and tracking the feature points. Secondly, we group the individuals by spatial neighborhood restraint. Finally, the groups would be further divided by motion correlation over time. These two restraints work with each other effectively and generate groups with coherent motion. Tested on many videos of real-world pedestrian scenes, this method can be applied to variety of scenes with different crowd densities and perspective of videos. We also make detailed analysis of other segmentation algorithms, and point out the shortcomings of these algorithms. Then the results of the present algorithm and our algorithm are compared. The outcomes show that our method proposed in this paper can segment crowd more effectively to the spontaneous small groups.On the basis of sub group segmentation, this paper proposes a new method of abnormal behavior detection based on deep convolutional neural network. In the first place, we take these sub groups as the research objects and define three kinds of group behavior descriptors, which exhibit motion features such as group interaction, stability and coherence comprehensively. Using these data as the training data facilitates the extraction of the behavior characteristics from video images. Then we use these descriptors and other image appearance features as the input of the neural network for training, which contains several convolutional layers and pooling layers. At the same time, we use the method of manual annotation to label the training video clips with different words according to different behavior subjects,scenarios and events. According to certain rules, we define some combinations of labels as abnormal behaviors. Then we use the labeled data to tune the outcome of the neural network and get good results.When group abnormal behavior occurs, we often need to guide and evacuate the group according to the situation of other spots in the location. Therefore, to meet the needs of practical application in real time monitoring of crowded scenes, this paper combines the OpenMP and openCV to realize a parallel processing system of multi-channel videos. The system can set up multiple threads to process multi-channel surveillance videos, so that we can achieve an omnidirectional monitoring of the same scene from different perspectives.