A Research Of Video-based Vehicle Behavior Recognition

In recent years,with the rapid development of urban traffic and improvement of computer performance,urban traffic has become more and more intelligent.The detection and tracking of video targets,as two important parts of the field of computer vision,play an irreplaceable role in the management of intelligent transportation.In the regulation of vehicle irregularities,violation of the direction of diversion and occupancy of emergency lanes is mainly achieved through human supervision.Human supervision has the problem of low efficiency and serious waste of human resources.To detect vehicle violation behavior intelligently,we should detect the vehicle and track it,then identify the vehicle behavior.Vehicle behavior recognition can be widely used in expressway,bridge,tunnel and other scenes to detect illegal behaviors.What's more,it can improve the efficiency of traffic police and reduce traffic accidents.This thesis mainly studies the vehicle detection algorithm under the surveillance video,vehicle tracking algorithm,multi-view vehicle behavior recognition algorithm.The main contents are as follows:We proposed Adaptive Selective Foreground Extraction(ASFE),which combines the frame difference method and the morphology.ASFE is proposed for the target with intermittent motion.It solves the problem that conventional foreground extraction algorithms cannot detect stationary targets.It solves the problem that the conventional foreground extraction algorithm cannot detect the stationary targets by two different background model updating modes.At the same time,ASFE improves the foreground extraction effect by morphological processing and the module of shadow removal.ASFE has achieved good results in detecting the target with intermittent movement.We studied a metric based on target center and candidate box.To solve the problem of tracking box offset in target tracking,in this thesis,a matching algorithm based on target detection is used instead of the target tracking algorithm,and the target track is obtained.In order to improve the performance of the target matching algorithm,this thesis proposes a metric based on the target center and candidate box.This metric combines the detected target candidate box and the target center coordinates,and reduces the problem of the remote target with small value and the near target with large value.And it achieves good results in the matching process of vehicles.We discuss a vehicle behavior recognition algorithm based on camera calibration.According to the different video shooting angle,the similar vehicle trajectory is different in different video.In this thesis,the camera calibration is added to the vehicle behavior recognition.The affine transformation is used to project the trajectory of vehicle,which reduces the influence of shooting angle.At the same time,we design a feature to record the behavior information of the vehicle,and classify the behavior by the feature.