Research And Implementaion On The Crowd Behavior Analysis Algorithms

Computer vision and intelligent analysis technology makes video analysis advanced, analyzing the behavior of people in video has become a popular research topic. Crowd behavior analysis can be applied to a variety of scenarios, such as: crowd management of people flow and estimation of crowd chaos in large activities, reducing congestion and preventing stampede. Smart video surveillance using computer vision technology can extract key information from videos, automatic process information and improve the efficiency of monitoring system, greatly reducing the workloads of monitor operators.Traditional surveillance video processing methods include background extraction, object segmentation, motion estimation, object tracking, face recognition, pattern recognition, behavior understanding and so on. Crowd video event analysis methods can be divided into two types: using feature of groups or using features of individuals. Methods using features of individuals are based on the segmentation and tracking of each target, and study crowd events with individual characteristics. Method using features of groups process and sample global image to obtain overall information, and analyze collected data to detect abnormal and normal behaviors. This approach focuses on the consistency of crowd behavior, so this method performs better in crowd analysis. The main content of this article is the analysis of crowd behavior in video, including the following aspects:1) Extend the social force model to rotation social force model. Social force model is derived from the crowd psychology. It can accurately model the motion state of crowds for better crowd modeling. Traditional social force model does not consider the rotation of the pedestrians, and this paper introduces the rotational torque produced by the rotational social force, models angular velocity and torque of pedestrians using physics modeling.2) Abnormal event detection based on rotational social force model. This article builds an abnormal detection framework based on rotational social forces. First, extract rotational features from crowds, and then train the crowd model using machine learning, and finally detect abnormal events with trained model. Experimental results show that this method can effectively detect abnormal crowd behavior, but it is difficult to detect abnormal behaviors of individuals.3) Cell event detection based on mesoscopic network model. This article builds a network model for both cell groups and cell individuals, estimates similarity of groups in two consecutive frames, and establishes inter group relations model, to detect the cells appear, disappear, move, split and merge events.