Research And Application Of Traffic Abnormal Behavior Detection Based On Deep Learning

Based on efficient transport abnormal behavior detection to ensure road safety has broad application prospects.The occurrence of abnormal traffic behavior is mostly attributed to the driver's abnormal behavior and the secondary accidents caused by the accident.This article starts from the two aspects of the driver's abnormal behavior and the road traffic abnormal behavior,and focuses on the driver's fatigue driving behavior and the road in the driver's abnormal behavior.Research on abnormal traffic behavior detection.In response to the intrusive characteristics and insufficient precision of the existing fatigue driving behavior detection methods,this thesis proposes a Pseudo-3D(P3D)convolutional neural networks(CNN)combined with attention mechanism to detect driving fatigue method.This thesis uses the structure of P3 D to merge the dual-channel attention module and adaptive spatial attention to construct a P3D-Attention network to remove the interference of background and noise on recognition;on the test set of the two data sets,the F1-score of the method in this thesis reaches 0.9989 and 0.9964,compared with Inception-V3 fusion LSTM method,F1 increased by 2%.In response to the problem of abnormal behavior detection in road scenes,a road traffic abnormal behavior detection method based on abnormal target detection and start time estimation is proposed.For the problem of difficult detection of small targets in traffic scenes,in order to ensure the recall rate of small targets in the distance,an abnormal vehicle detection method based on the perspective view module combined with YOLOv3 is proposed.The single target tracking result is used to obtain a more relevant target to solve a better perspective relationship diagram,and the YOLOv3 target detection method is used to detect redundant and cropped regional images of appropriate capacity,and the abnormal vehicle detection is performed quickly.The detection reaches 0.97 in the relevant data set.F1 score.Regarding the shortcomings of existing methods for estimating the start time of abnormal traffic behaviors,which rely too much on advanced trajectory features,are expensive and time-consuming,a method for estimating abnormal traffic time based on the siamese cross-correlation mechanism and P3D-Attention network is proposed.The siamese cross-correlation mechanism makes the network pay more attention to the selected target,model the spatio-temporal characteristics of key frames to improve the prediction performance of abnormal traffic vehicles,solve the problem of abnormal traffic time detection,and accurately estimate the start time of abnormal behaviors on the relevant data set.The root mean square error is 9.22,and the execution time of each video at speeds of 983 frames per second,the average execution time of every 12 frames video is only 12.2ms,compared with the use of position detection box and tracking trajectory method,the consumption is smaller and the time consumption is shorter,the consumption is smaller and the time-consuming is shorter.On this basis,this thesis designs and initially implements an abnormal traffic behavior detection system based on deep learning.It provides the relevant traffic management departments with methods for detecting driver fatigue behavior,detecting abnormal traffic targets,and estimating traffic time for abnormal traffic behaviors.A new and efficient method for detecting abnormal traffic behavior.