Design And Implementation Of Multi-scene Traffic Video Object Detection System

In recent years,more and more attention has been paid to machine vision.As one of the most important branches of machine vision,image processing and object detection technology has been widely used in various fields,such as unmanned vehicle,security monitoring,aerospace and so on.Based on the application background of unmanned vehicle and intelligent transportation,for traffic video scences,this paper tests the targets of structured road lines,pedestrians and vehicles,and urban road sand table simulation vehicles.The main research contents are as follows:(1)Design and implementation of lane detection system for driving video.In order to detect the lane line of the driving video in real time,the system can cope with the highway scene and ensure the accuracy and stability of the detection,This paper designs a lane line area search method.Specifically,The first step is to set the ROI region of interest,extract the pavement area and increase the detection speed;The second step is to enhance the edge features in the vertical direction of the lane line,eliminate image noise,optimize the detection effect of Hough line transformation,and improve the accuracy of detection.;The next step is to perform a sub-regional search on the detected line model,identify the left and right lane lines and assign label information;According to the label information of lane line and the detection results of the previous period and the current period,the fitting update mechanism is designed to improve the stability of detection.Finally,according to the detected lane line,the driving area is delimited and Output vehicle lane change information.The experimental results show that the accuracy of the lane detection is 97.2% and the time of each frame is 33 ms,it can meet the requirements of stability and real-time in the field of unmanned driving.(2)Design and implementation of target detection system.In order to improve the accuracy and real-time performance of object detection in the intersection monitoring scenes and urban roads sandbox scene,This paper implements a target detection system based on YOLOv3,which solves the problems of light changes,small object,background complexity and object occlusion in the scene.Specifically,the FPN structure is used to realize multi-scale feature extraction,The characteristics of shallow network and deep network are more fully integrated.By presetting nine scale priori boxes and the object confidence elimination mechanism,the object position and category can be predicted more quickly and accurately.At the same time,in order to improve the generalization ability and robustness of the model,the data set is processed with the data enhancement method.The experimental results show that the detection accuracy of pedestrian and vehicle under the intersection monitoring is 96.9% and 94.3%,and the time of each frame is 83.3ms.The detection accuracy of the intelligent car,uav,cross-country car and disc car is 93.2%,95.3%,90.5% and 96.2%.The time of each frame is 83.3ms,It can meet the requirements of intersection monitoring and UAV aerial video analysis accuracy and real-time performance.