Research On Lane Detecting And Road Sensing Towards Driverless

As an important part of road environment perception system,lane detection is the basis of key driving technologies such as lane departure warning and lane keeping,and it is of great significance in autonomous driving technology.Based on the demand of lane detection in highway scenes,this thesis implements a fast lane detection method with both high accuracy and high real-time by improving image pre-processing and optimizing the network structure of the detection model.On this basis,a multi-task road environment perception method including lane detection,vehicle target recognition,and driveable area segmentation is proposed by improving the road perception model.The main research of this thesis is as follows:1)The proposed road image grid division pre-processing method improves the speed and accuracy stability of lane detection.In this thesis,we propose a road image grid segmentation method and adjust the images to different resolutions to train separately.Experiments show that the method can achieve 95.97% accuracy and 400 FPS in Tusimple dataset,which is 0.15% and 30% better than UFLD respectively.The detection accuracy of most scenes in Cu Lane dataset is better than UFLD algorithm,and the average accuracy is improved by 1.7%.2)The Shuffle-Inception structure is proposed to improve the detection accuracy of the lane detection model.In this paper,we combine the advantages of Inception module and Shuffle Net network structure,and propose Shuffle-Inception structure with the help of SE Attention module to improve the performance of lane feature extraction.Experiments show that in the Tusimple dataset,the improved model improves the accuracy by 0.34% compared with UFLD.In the Cu Lane dataset,it outperforms the original method in both conventional road conditions and curve detection scenarios compared with other open source lane detection algorithms,with detection accuracies of92.3% and 70.2%.3)Proposed a multi-task road environment sensing method based on improved YOLOP.In order to improve the detection capability of small targets,the thesis uses Shuffle-Inception to extract features and Anchor thickening to improve the detection capability of small targets.The null convolution fusion module is designed to reduce the proportion of missed targets in area detection.For lane detection,the location information of features is more concerned,and the lane is more concerned with the location information of the features,the lane branches are induced separately and an improved APN structure is added to improve the detection capability of lane detection.The experimental results show that the method can effectively solve the problems existing in YOLOP,while the lane line detection accuracy reaches 72.5%;the detection speed reaches 54.6 FPS,which is improved by 12.6%.