Research On Video-based Detection Method Of Several Kinds Of Expressway Traffic Incidents

By the end of 2021,the mileage of expressways in our country has reached 169100 kilometers.Due to the fast driving speed and high closure of expressways,traffic incidents such as pedestrian entry and illegal parking are liable to cause serious accidents.Therefore,timely detection of the according incidents is of great significance for improving the operating efficiency and safety of roads.However,the existing manual monitoring method consumes large human resources and prone to missed detection.Thus,based on road monitoring video and deep learning technology,this thesis proposes a traffic incident detection algorithm by cascading object detection,tracking and incident recognition models to realize real-time detection of pedestrian,parking and congestion incidents which are characterized by high frequency and great harm.Besides,this dissertation develops a corresponding traffic incident detection system.The main contents of this dissertation are summarized as follow:(1)In view of the problem that the benchmark YOLOv5s model has low accuracy due to missing detection of pedestrian and distant vehicle targets,a YOLOv5s-ST object detection algorithm suitable for expressway is proposed.Firstly,this dissertation constructs the backbone network based on multi-head self-attention mechanism to enhance the feature extraction ability of the model for targets.Furthermore,this thesis designs a multi-scale feature fusion network to make full use of high-resolution feature maps and original feature information,which can reduce the missed detection of small targets.The proposed improvements meet the real-time requirements of detection on self-built expressway object detection dataset,and achieve higher detection accuracy.(2)Aiming at the problem of frequent vehicle ID switching of the benchmark DeepSort algorithm during the tracking process,a DeepSort-occ multi-object tracking algorithm suitable for expressway is proposed.A ResNet-DS network with better performance is designed by adding residual structure and removing redundant fully connected layer in the algorithm.Besides,center loss is combined to constrain the distance between features and their corresponding feature centers,which enables the network to extract more discriminative features and exerts the advantage of apparent metric of DeepSort algorithm.The proposed improvements meet the real-time requirements on self-built expressway multi-object tracking dataset,and have a more stable tracking effect.(3)Aiming at the problem of impractical speed calibration acquisition process and difficulty in determining the optimal discrimination threshold for indirect indicators of videobased congestion recognition algorithm,a multifactor decision making congestion recognition algorithm is proposed.This thesis defines several indicators such as the ratio of parking vehicles Rt and the ratio of slow-moving vehicles Rlafter analysis,which indirectly reflect the road status.On this basis,the above indicators are inputted into support vector machine to recognize the congestion on road.The experimental results on the self-built traffic incident recognition testing video set demonstrate the feasibility of the proposed algorithm.(4)A corresponding expressway traffic incident detection system is designed and realized.This thesis deploys the above traffic incident detection algorithm in the server environment and implements the related business logic.Besides,a web app is developed to complete interaction and result display tasks.In addition to the core function of incident detection,the implemented system also provides auxiliary functions such as alarm notification and data statistics.