Traffic Scene Understanding Based On Monocular Vision

With the development of unmanned driving technology,intelligent vehicle and intelligent detection system cooperate to gradually replace human beings to obtain driving environment information and make decision judgment.Advanced Driving Assistance System(ADAS)is an important prerequisite of unmanned driving technology.With the development of unmanned driving technology,this passive warning role gradually turns into active decision intervention,so ADAS technology.The improvement of technology is an important prerequisite for the safety,advancement and stability of unmanned driving technology.This paper designs a traffic perception system based on monocular vision,which provides a system scheme for ADAS system to process the collected video information.Generally,visual driving assistance system is mainly based on computer vision.Sensors are divided into monocular and binocular cameras.Compared with monocular cameras,the cost of monocular cameras is lower and the computational resources are less.Therefore,monocular cameras are preferred as the technical route in this paper.The design framework of this traffic sensing system is as follows:Firstly,this paper designs a vision-based Kalman filter lane detection and tracking system.Through video image preprocessing,sliding windows are used to extract the feature information of left and right lanes,and these feature points are fitted twice to draw a complete Lane line.At the same time,in order to improve the robustness and accuracy of the algorithm,Kalman filtering algorithm is added.When the road information is incomplete,special driving conditions and road disturbance information are more,the lane line can still be extracted completely,with the accuracy of 74.8% and the speed of 332 frames per second.Secondly,a target detection algorithm based on depth learning is designed to identify the vehicle types and image location information around the body.By studying and comparing the performance of various algorithms,MASK R-CNN is chosen as the target detection framework,MS COCO and KITTI data sets are used as training data sets,and model files are obtained after adjusting the hyper-parameters.The model is used to test the video,and the final detection results are obtained.The average recall rate is 94.58%,and the accuracy rate is 92.76%.Then an algorithm for target tracking is designed based on the position information detected in the image.The location information of the target detected before is used to track the surrounding motor vehicles,non-motor vehicles and pedestrians,so as to provide the basis for calculating the relative distance of the vehicle,and to provide assistance for collision warning or other auxiliary driving.Combined with the target detection system,the test accuracy reaches 70.7%,and the speed is 25 frames per second.Finally,the lane detection,target detection and tracking algorithm are fused,and a traffic scene perception system for multi-target detection and tracking is designed.Ubuntu is used as the operating system to test and verify each part.The simulation test proves that the system can detect and track the target in the traffic scene in real time,and the detection accuracy achieves the feasibility and real-time performance of the system.