Detection And Cross-mirror Recognition Of Brawls Inside And Outside The City Bus Compartment

In recent years,the sudden abnormal behavior of groups in the bus has attracted gradually more and more attention from the whole society.Bus safety maintenance has become an important public safety issue that needs to be solved in all countries in the world.The internal conditions of public transport vehicles are complicated,and when conflicts arise between passengers,it may be accompanied by fights between passengers.Therefore,the use of video surveillance system to alarm the brawl behavior of people in the bus compartment and transmirror identification of the identity of the people in the bus can ensure effectively the safety of vehicles and personnel,while preventing more serious consequences.Based on this,this thesis combines image processing,deep learning and video analysis to study the brawl detection and transmirror recognition methods in the bus car scene,and the main work content is as follows:(1)The construction of brawl re-identification dataset,taking the brawl behavior that has a large impact on the life safety of personnel in the bus compartment scenario as the construction object of the dataset,the data collection method,some environments and equipment collected in the field and the process of data processing are mainly introduced,and a brawl re-identification dataset in the bus compartment scenario is constructed finally,which provides data support for the training and testing of the subsequent brawl detection algorithm and brawl personnel re-identification algorithm.(2)The design of brawl behavior detection algorithms,a brawl behavior detection algorithm combining object detection and video classification is proposed,which can be applied to the interior of the bus compartment,which can solve the problems of serious occlusion caused by human interference and difficult acquisition of personnel features.Among them,the object detection algorithm is used to obtain the position of the human target bounding box,and the video classification algorithm is used to obtain the position of the human target bounding box where limbs overlap in object detection and extract its features,and then combine the two to detect the brawl behavior in the video.For the experimental process,data sets,including human target data set and vehicle personnel fight data set,were collected and made for the training and verification of the algorithm.The experimental results obtained by dividing the experimental scenes and testing show that the designed brawl behavior detection algorithm can overcome some human interference caused by brawlers in the carriage,so as to detect accurately brawling behavior.(3)The design of brawler re-identification algorithms,based on the previous brawl behavior detection algorithm,combined with the facial characteristics of personnel,a transmirror recognition algorithm for brawlers based on target detection,target tracking and pedestrian rerecognition network is designed,which realizes the transmirror recognition of brawlers in car inside and outside the carriage.During the experiment,the experimental scene is divided and the designed transmirror recognition algorithm is trained and verified through the self-made data set,and the detected experimental results show that the algorithm can overcome the change of viewing angle and light inside and outside the carriage when the occlusion of the brawlers is not very serious or there is no occlusion,so as to identify accurately the brawlers across the mirrors.