A Study On Vehicle And Lane Detection Based On Machine Vision

The rapid development of economics has brought about the fast popularization of vehicles.While greatly improving the transportation,automobiles also cause some urgent social issues related to resources,environment,safety and traffic,which subsequently drives the need for unmanned driving technology of the whole society.Nowadays,unmanned driving technology has become a hot issue among universities.This paper mainly focuses on the environmental perception which is an important link of the unmanned driving technology.Detection algorithms have been successfully designed to detect surrounding vehicles during the self-driving and lane lines of the structured roads.On the basis that DCNN(deep convolutional neural networks)can learn the vehicle characteristics adaptively,the detection of vehicle targets in driving scenes was realized via an end-to-end real-time target detection system(YOLOv2).First of all,data of targets to be detected were collected,and the weights of the model parameters were acquired through a pre-training by Darknet-19.Then the vehicle detection model was trained by YOLOv2 network using the vehicle data set labelled in the actual traffic scene diagram.The optimal anchor depth-width ratios of vehicles to be detected were obtained by the mean clustering method related to IOU deviations.And the optimal prediction block diagram was obtained by NMS(non-maximum suppression).At the last layer of YOLOv2,the location of the vehicle target would be predicted directly.After several iterations,a vehicle detection model with better parameter weights was eventually obtained.Through the simulation tests against actual vehicles and the real road scenes,the model was proved to quickly and accurately detect the vehicle targets in the scene under the influence of light reflection,small occlusion and different visual angles.As for the detection of lane lines,traditional image processing techniques were applied.Firstly,the algorithm flow for the extraction of the optimal image edge information was proposed:grayscale processing by weighted averages method,median filtering noise reduction,partial adaptive threshold binarization and Canny edge detection algorithm.An improved weighted clustering method was proposed to detect the vanishing point,which can effectively extract the region of interest and reduce the loss of useful information.The edge noise was then filtered twice by assigning the weight coefficient of the neighbourhood position of the pixels.Finally,the least square method was used to complete the lane line fitting.The results of simulations against the real structured roads showed that the lane detection algorithm proposed in this study can stably detect the lanes even with the influence of shadows and road interference.With the integration of the detection algorithms of vehicles and lane lines,the prediction of driving environments can be realized.