The Research On Detection And Recognition Of Abnormal Events In Traffic Intersection

Abnormal event in traffic intersections is an important affecting factor of traffic operation.As a high connection of public travelling,the traffic intersection has w wide range of impacts.With the gradual improvement of urban development and living standards and the rapid urbanization process,Beijing has rapidly become the world’s top 20 mega-cities.With the development of the city and the rapid increase of the number of motor vehicles,the urban traffic environment deteriorates rapidly,and traffic congestion becomes the normal state of the city.According to the statistics,60-70%of traffic congestion is affected by traffic accidents.This research provides research ideas for the detection and identification of abnormal events at traffic intersections and provides technical support for the identification of abnormal events at traffic intersections.It is of great significance for the timely discovery and processing of abnormal events at traffic intersections to reduce or relieve traffic congestion.It can improve the timely feedback and rescue efficiency of related departments for abnormal events and avoid the expansion of hazardous events.Also,it can effectively reduce the harm to the safety of people’s lives and property.As a fast,nondestructive and efficient detection method,video-based event detection has become the main means of traffic events monitoring.Due to the complex traffic environment and diverse traffic operating conditions,the video-based automated detection method is not perfect,mainly existing as follows:1)The detection of most traffic abnormal events can only detect the occurrence of abnormal events.However,it cannot identify the type of traffic abnormal events.Thus,it is difficult to realize the rapid processing and feedback of abnormal events.2)Existing abnormal traffic events generally focus on the operation of traffic vehicles while ignoring the weak and dominant position of pedestrians in abnormal traffic events.3)With the extensive application and continuous development of monitoring equipment,traffic monitoring data has increased dramatically.The traffic event detection based on video cannot fully realize automatic processing.People usually check the monitoring video artificially after the occurrence of abnormal events.This increases human cost and wastes time.As a self-selected subject,this paper studies the types of traffic intersection abnormal events and related technologies in terms of the impact of traffic intersection abnormal events.The paper mainly includes the following 3 aspects:1)The accurate segmentation method of pedestrian target is applied to the accurate segmentation of individual pedestrian target.2)Low rank texture description of human behavior used for group anomaly event type recognition.3)Human posture estimation based on human key point analysis is used for pedestrian falldown event type recognition.The main research work and innovation points of this paper are as follows:(1)In view of the influence of background information within the detection range of traffic intersection target,a two-window Fisher adaptive target segmentation method based on global significance is proposed which can quickly and accurately extract the target and reduce the influence of the target background.Aiming at the problem of relative continuity and unity of the neighborhood environment of individual target,this paper proposed to extract the target’s salience graph by global significance.Combined with double threshold segmentation window has simple structure and can balance the error of different local information and Fisher linear threshold can accurate judgment reasonable selection of adaptive threshold.The proposed double window Fisher adaptive threshold segmentation method could significant figure of individual target threshold segmentation,achieve the accuracy of the individual target extraction and avoid the influence of individual target neighborhood background information(2)Aiming at the complexity of population density study,a human body description method based on low-rank texture direction is proposed.Based on it,a population hetero kinetic event model is constructed to identify different population hetero kinetic events.The texture features of the human body itself change with the change of human behavior,and the human body target is extracted as the overall texture.As the human body structure driven by the direction of human action has a tendency,it is shown in the image that the selected individual target area is a sparse matrix area,with a high rank.When the human body is upright or relatively upright,the matrix of human body area is dense and also has a higher rank.The difference is that sparse matrices can be compressed to obtain a lower matrix rank.The dense matrix does not decrease the rank of the matrix as the matrix is compressed.Low-rank texture feature extraction can compress selected image regions to obtain a lower matrix rank.The low-rank direction of the selected image region is obtained by calculation.Then the image selection region is transformed to form the low-rank texture direction and motion trend description feature of the image target.The group abnormal events are the overall characteristics reflected by the mutual relations among multiple objects.The relationship between the direction and motion trend of multiple objects is constructed to form the detection model of group abnormal events.(3)In view of the complexity of the cause of pedestrian falling at traffic intersections,a posture evaluation method based on the analysis of key points of human bones is proposed.Based on this,a pedestrian falling event recognition model at intersections is constructed.The change of human posture is accompanied by the transformation of the relationship between the key points of human skeleton.With the key points of human bone neck as the center,8 key points(left and right shoulder,right elbow,left and right hip,right and left knee)that are relatively important in human movement are selected.The key position information is extracted to build the relationship between the key points.Taking the key points in the neck as the reference point,we can calculate the entropy value of eight distances from the reference point and the ratio of the distance entropy of four key points from the right elbow and left knee to the reference point and the distance from the right shoulder and left hip of the middle key points to the reference point so as to build the posture evaluation model of human body target.If the human body is in an abnormal state,the target is taken as the center point,the target within the neighborhood range is searched,the Euclidean distance between the target and the central target is calculated,the relationship between the neighborhood target and the central target is determined,and the type of ground fall event is realized.In this paper,the traffic intersection abnormal events recognition and related technologies are explored and studied,but there are still the following deficiencies to be further studied:1)For the monitoring video obtained on the basis of the existing monitoring system,a large part of video content cannot be detected and identified using the Faster-RCNN model.Especially when the monitoring video is fuzzy or the monitoring target is small,the recognition is more difficult,which seriously affects the detection and recognition results of the target.2)The low-rank texture orientation of the target is affected by the overall target as well as the target’s own texture,such as clothes,backpacks,and lighting.How to further select and use the low-rank texture direction of human body and describe human behavior becomes the focus of research.3)At present,this paper only identifies and analyzes the traffic intersection group abnormal events and pedestrian fall events.It is necessary to identify and expand other types of traffic intersection abnormal events.