Research On Vehicle And Lane Detection Method Based On Convolutional Neural Network

Nowadays,with the development of science and technology,artificial intelligence has gradually spread to every corner of our lives.As a hot topic in the field of artificial intelligence,autonomous driving technology has also attracted more and more researchers.attention.Among the many reference factors of autonomous driving technology,vehicles and lanes are important reference objects during driving.Vehicle detection is the most basic and important part of autonomous driving technology.The ability to timely and accurately detect and identify vehicles ahead is the key to ensuring the safety of autonomous driving.Lane detection is a ruler to prevent vehicles from driving away from a reasonable area,also plays an important role.Vehicle detection and lane detection both belong to object recognition and object segmentation problems in the field of image processing.With the rise of the deep learning boom,related technical theories of convolutional neural networks have been increasingly applied to such problems.This paper takes this as a foothold and launches a research on vehicle and lane detection methods based on convolutional neural networks.The specific content is as follows:Aiming at the problem of vehicle detection,the research of vehicle detection method based on multi-task convolutional neural network was carried out.Combined two open-source vehicle data sets and self-collected data sets to produce a hybrid experimental data set;a multi-task convolutional neural network model was built,and its internal structure was cascaded to improve the computing speed.Finally,the trained model was put into the experiment,and the overall test accuracy reached 95.0%.Aiming at the problem of lane detection,a research on lane detection method based on full convolutional network was carried out.After reclassifying and labeling the unclear lanes in the sample image,a lane dataset detection data set was constructed;a full convolutional network model was built,and the network structure was adjusted to improve the accuracy of detection.Finally,the trained model is put into experiments,and lane images under various scenes are detected and identified.The overall accuracy rate is 95.2%.At the same time,a comparison experiment with the classic method Hough transform has been carried out,and better results have been achieved.In the last chapter of this paper,research on joint detection methods for vehicles and lanes was carried out.The joint training network model was constructed by setting up multi-task training in the volume neural network;the VGG16 model was improved and improved to optimize the performance of the joint detection network;and the feasibility of model training was obtained by formulating special model training strategies Finally,the trained model was put into the experiment,and the overall accuracy of the test on the KITTI dataset reached 89.2%.