Research On Highway Traffic Incident Detection Based On Deep Learning

Up to now,the total mileage of expressways in China has reached 161,000 kilometers,and the world’s largest expressway network has been built.Although this has brought convenience and speed to people,it has also caused a series of serious traffic problems.The location of the highway is relatively remote and the speed of the vehicle is fast,and once a traffic incident occurs,it is often more serious.The traditional manual monitoring method is still the main method of highway traffic incident detection.The manual monitoring method is expensive and difficult to detect the occurrence of the incident quickly.In this paper,the deep learning detection algorithm is used to conduct in-depth research on highway traffic incident detection.In order to improve the real-time and accuracy of incident detection,a highway traffic incident detection model is established.Classify and define highway traffic incidents,define pedestrian break-ins and traffic accidents as traffic accidents,and define reverse driving and speeding as traffic violations.By collecting highway surveillance video and picture data in multiple scenes and different environments,the video images are screened and processed,and a data set of vehicles,pedestrians,and traffic accidents in the format required by the detection model is established to provide data support for the establishment of the detection model.In order to meet the real-time and high-accuracy requirements of highway traffic incident detection,this paper uses a deep learning-based target detection algorithm to design a highway traffic incident detection model.According to the factors affecting the performance of the model,an improved strategy for traffic accident detection algorithm based on deep learning YOLOv4 is proposed from four aspects: anchor frame mechanism,multi-scale feature layer,network lightweight and non-maximum suppression.The K-means++ algorithm is used to re-cluster the model to improve the convergence speed of the model;due to the fixed camera on the highway,the size of the vehicle,traffic accidents and pedestrian objects will change.Select and integrate feature maps of various scales for prediction to improve the detection ability of highway traffic accidents;in order to meet the real-time requirements of highway traffic event detection,depthwise separable convolution is used instead of standard convolution;in order to prevent missed detection due to overlapping targets,NMS is replaced by Soft NMS.Finally,through experiments to compare the detection effect of the network model before and after the improvement on traffic accidents,it is determined that the improved detection model of YOLOv4 is used to detect traffic accidents and vehicles.Considering to meet the real-time requirements of traffic violation detection on actual highways,this paper proposes to improve the apparent feature extraction method based on the Deep SORT multi-target tracking algorithm to the HOG feature extraction method,and combine the improved deep target detection network YOLOv4 with the target tracking algorithm.Track the vehicles in the highway surveillance video and extract the vehicle trajectory features.Through the vehicle trajectory features combined with the specific algorithm of the violation events,the vehicle reverse and speeding traffic violation events appearing in the video are discriminated.This paper aims to realize the detection of highway traffic accidents（car accidents,pedestrians）and traffic violations（reverse driving,speeding）,and to improve the real-time and accuracy of highway traffic incident detection,establish a highway traffic incident detection model.The experimental results show that the model established in this paper can meet the requirements of high real-time and accuracy of event detection in highway traffic scene.