Scene Perception Guided Crowd Anomaly Detection

With the increase of social population and the enhancement of human safety awareness,intelligent monitoring has gradually become an important demand in today's society.At present,intelligent monitoring has been widely used in shopping malls,schools,stations and other places,playing a key role in the field of public safety.As an important branch of intelligent monitoring,crowd abnormal behavior detection is closely related to public safety.Crowd abnormal behavior detection technology can automatically analyze crowd behavior in video scenes and give real-time alarms to abnormal events,thereby reducing the occurrence of safety accidents,and has great research value and application value.For crowd abnormal behavior detection,the accuracy of abnormal behavior detection mainly depends on the division of regions of interest and feature extraction.However,the region of interest used in the existing methods is usually a rectangular block of a fixed size,which destroys the integrity and consistency of pedestrians in the of interest,resulting in missed detection.In addition,the features extracted by such methods cannot effectively characterize crowd movement,leading to false detections.In order to deal with these issues,we put forward a approach for detecting crowd abnormal behavior based on scene perception and fluid forces.The specific content and innovation of the algorithm are as follows:(1)Aiming at the problems of traditional regions of interest,an adaptive size consistency group is used to replace fixed-size regions of interest.In order to reasonably use the spatial information and attribute information of the motion crowd,the scene perception theory is used to divide the motion crowd.(2)Based on the separation processing mechanism in the scene perception theory,a divide and conquer clustering algorithm is proposed.The algorithm can make use of the spatial information and attribute information of the motion crowd reasonably,and the steps include: connected region screening,spatial information clustering and attribute information clustering.In addition,the method of noise groups merging is used to eliminate noise interference,thereby ensuring the integrity of the consistency group.(3)Based on the smooth particle hydrodynamic,a method for characterizing crowd movements using fluid features is proposed.The motion crowd is regarded as a group of interacting fluid particles.The mass force and surface force of the fluid particles in motion are calculated as the fluid features of the consistency group.Fluid features can greatly describe the crowd movement,so the abnormal behavior and normal behavior can be effectively distinguished.The research findings prove that the proposed crowd abnormal behavior detection method based on scene perception achieves higher accuracy in comparison with some existing methods in terms of both frame-level and pixel-level measurements.