Research On Campus Fighting Behavior Detection Based On Skeleton Sequences

In recent years,school bullying has happened from time to time,which has aroused public attention.School bullying may cause great harm to victims physically and mentally.However,for reasons such as self-esteem and fear of retaliation,the victims often do not take the initiative to report the incident to teachers and parents.As an important means of security prevention,video surveillance has developed rapidly in the past decade.The campus is basically covered with surveillance cameras as well.However,it is unrealistic for security personnel to stare at surveillance video continuously for a long time.The video is often played back for verification after the accident and it is difficult to intervene in school violence in time.Therefore,this thesis studies the identification of violence based on campus surveillance video,which allows the school to use the existing campus monitoring system to detect and intervene in campus violence in a timely manner,protect the victimized students,and make a useful supplement to campus safety without the purchasing or installing professional equipment.The research contents of this paper mainly include the following aspects:(1)Research on multi-target tracking based on skeleton extraction.Skeleton is a stable and robust human body feature.In this paper,Open Pose is used to extract the bone joint points of each frame in the surveillance video and cluster the human body skeleton.After obtaining the set of all human body skeletons,multi-target The tracking problem is transformed into a probabilistic model that divides the skeleton set.Among them,the skeletons in each subset are extracted from the same person in different frames.In this paper,this problem is solved based on the Markov chain Monte Carlo data association method,and the skeleton between frames is correlated to achieve more accurate multi-target tracking.(2)Research on Recognition of Campus Violent Behavior Based on Skeleton Sequence.This paper is based on the ST-GCN behavior recognition algorithm to detect whether there are violent behaviors in campus surveillance videos.The skeleton extracted from the video is first constructed as a spatio-temporal diagram of the human skeleton joint sequences,so that the diagram can simultaneously reflect the time information and spatial information of the skeleton joint points.Then the convolution method of the two-dimensional image is extended to the convolution of the skeleton spatio-temporal graph,and the attention mechanism is introduced to assign different weights to the joint points.Use the Fight-skeleton dataset constructed in this paper to train and test the network model to realize the recognition of campus violence.(3)Designed and implemented a prototype system for identifying violent behaviors based on campus surveillance video.The system is mainly composed of two modules: multi-target tracking and violent behavior detection.Using the public data set RWF-2000 for verification,the method used in this article can reach an accuracy of 87.75%.