Resarch On Lane Detection Based On Computer Vision

With the development of artificial intelligence,autonomous driving and ADWS system have become the current research hotspots,in which lane detection plays an important role.In real world,lane detection can be very challenging due to the variety of nature environment and road scene.Not only accuracy but also real-time performance is required in lane detection,to help the autonomous driving or ADWS system response in time.In this thesis,with the use of Convolution Neural Network,we propose an algorithm called FastLane which can detect an unfixed number of lanes accurately and in-time,which mainly includes the following works:(1)For the phenomenon that with the increment of distance,the worse performance of lane detection caused by perspective transform,an inverse perspective transform pre-processing has been introduced and realized.By transforming the image into a bird-eye view,semantic segmentation plus simple post-processing can be used instead of time-consuming instance segmentation or clustering post-processing,improving algorithm speed.(2)Aiming at the visual characteristics of lanes,a multi-scale real-time semantic segmentation network has been proposed.The network applys dilation convolution and multi-scale features,as well as combines the underlying features to obtain accurate segmentation results under the guidance of multi-scale semantic features.Also,the network structure is optimized to improve the running speed.(3)Combining the previous inverse perspective transform transform and semantic segmentation network,a lane detection algorithm called FastLane is proposed.The FastLane algorithm firstly uses the self-learning inverse perspective transform,and then uses the binary semantic segmentation network to segment the image,and finally post-processing.In the actual road situation and open source dataset TuSimple,the thesis makes a full experimental analysis and comparative experiment on the FastLane algorithm,which verifies the effectiveness and strength of the algorithm.Compared with other classic SOTA algorithms,FastLane algorithm can not only realize real-time lane detection,but also detect any number of lanes,which has practical significance in traffic scenarios.