Research On Vehicle Object Detection And Application Based On Deep Learning In Expressway Scenes

Traffic statistics and behavior analysis of vehicle objects are of great significance for traffic management in road communication.In Intelligent Traffic Video Analysis Technology,The detection of vehicle objects is usually the basis of other research work.Develop an efficient and robust detection algorithm that plays a key role in traffic video analysis systems.This article takes the highway traffic scene as the research object,Combining deep learning based object detection technology and target tracking technology,Realize accurate detection,classification,vehicle counting and abnormal event detection of vehicle object.This paper has fully studied,analyzed and compared the characteristics of the current major learning object detection frameworks(Faster R-CNN and SSD).The real-time and accurate SSD framework was selected as the basic structure of the object detection.For the highway scene,the format,method and rules of the sample annotation,Completed the data set labeling task for a total of 45,794 vehicle object including three types of vehicles(Car,Bus,Truck),three angles(front,side,rear),and two time periods(day and night).On this basis,a comparative experiment of the basic net model of SSD framework(Base Net),Determine VGG16 as a basic model;Using the self-made data set,the Fine-Tuning method is used to train and tune the SSD model,The training and tuning of the SSD model were performed to generate the deep learning based object detection model used in this study.Through this model,not only reliable and accurate detection of the target can be achieved,but also accurate vehicle classification can be achieved.A multi-object matching algorithm was designed,When the object is detected,the same object and different frames are associated with each other,In combination with the object tracking algorithm to obtain the target's trajectory,Track-based vehicle counts and detection of abnormal behaviors such as retrograde,non-stop,etc.are achieved.The algorithm in this paper has been tested and the detection accuracy of the vehicle object has reached 89.4%,The classification accuracy rate is 99.6%,the count accuracy rate reaches 97.5%,In a single NVIDIA 1080 Ti GPU server,The target detection processing speed reaches 13-22 fps,basically meeting the real-time requirements.