Lane Line Detection Based On The Encoder-decoder Network Structure Of LargeFOV

One of the key technologies in Advanced Driving Assistance Systems is lane line detection.In some simple scenarios,the performance of the existing lane detection algorithm can meet the actual application requirements,but the performance of the algorithm will drop sharply in the variability of road scenes such as poor lighting,crowded,lane line appearance changes,landmark interference,etc,the effect of existing lane line detection algorithm will drop dramatically.In recent years,deep learning based methods have pushed computer vision to a new level.It is far superior to traditional algorithms in terms of robustness and accuracy.Therefore,this paper attempts to apply deep learning to lane line detection,and achieve real-time and accurate results in in different road scenarios.This paper aims to explore a lane line detection network that can cope with changes in road scenes,considering the real-time and accuracy,combined with the actual application scenarios,the lane line detection algorithm LargeFOV-LaneNet based on LaneNet and the lane line detection algorithm Spatial-LargeFOV based on Spatial CNN are proposed.The main work and innovations of this article are as follows:(1)LargeFOV-LaneNet,a multi-task network,is proposed based on the basic model of LaneNet.LargeFOV-LaneNet explored the encoder-decoder structure based on LargeFOV,combined with the atrous convolution combination with different dilated rates and the skip connection,which performed the fusion of multi-scale contextual information.Experiments show that the lane line detection algorithm can realize real-time and accurate in different road scenarios.(2)Spatial-LargeFOV is proposed based on the basic model of Spatial CNN.Build an encoder-decoder network which applys SCNN to learn the spatial relationship of the lane line.Experiments show that this algorithm realizes real-time and accurate lane line detection in different road scenes and strikes a trade-off between speed and accuracy.The experimental results show that the network Spatial-LargeFOV proposed in this paper has good robustness in variably road scenes such as poor lighting,crowded,lane line appearance change and landmark interference.The evaluation result in CULane test data set can achieve the F-measure value of 71.5 and the speed of 35 ms per frame.