Research On Lane Line Detection Algorithm Based On Deep Learning

With the continuous maturity of artificial intelligence technology,unmanned driving technology has become a hot research direction,and lane line detection technology can effectively ensure the vehicle driving in the correct area in the driving assistance,which has important research significance.In this paper,the deep learning algorithm is mainly used to process the images of the surrounding road condition information of the driverless vehicle collected by the on-board camera,so as to realize the fast and accurate detection of the lane line.At present,the semantic segmentation method is mainly used to detect lane lines.However,semantic segmentation methods still have some problems.First of all,with the need of semantic segmentation of vehicle to collect the image of one by one,to extract the feature information of pixels are classified,and the lane line usually account for only a few pixels,waste a lot of computing resources,it is difficult to guarantee the lane line detection,real-time and through semantic segmentation method,can only be lane line with background information,unable to form different lane line as an example,to only to test the predefined number of lane line.Secondly,the existing lane line detection methods generally have a high false detection rate,and the safety of unmanned driving is difficult to be guaranteed.Therefore,this paper proposes a method to detect the lane line by using the improved hourglass module to generate key points.The lane line detection algorithm proposed in this paper mainly consists of two parts: the lane line detection network based on the improved hourglass module and the post processing algorithm for the high lane line false detection rate.The overall design of the network takes into account the detection accuracy and detection speed.This paper puts forward the improvement of the hourglass module adopts the structure of encoder and decoder and short,and two pieces of the hourglass stack module approach for image feature extraction,makes the whole network of local information and global information to extract more accurate,the last generation lane line and the generated key point is far less than the number of input image pixel size,greatly reducing the amount of calculation,improve the detection speed model.The lane line post processing algorithm mainly includes lane line clustering and lane line fitting and correction.The idea of clustering is used for instance segmentation to divide different lane lines into different instances,which effectively solves the problem that semantic segmentation method can only detect the pre-set number of lanes.Through the subsequent fitting correction algorithm,the wrong lane points generated by the model are filtered out,and the false detection rate of lane lines is reduced.The experimental results show that compared with the semantic segmentation method,the proposed algorithm reduces the computation amount in feature extraction,and has a faster detection speed,the fastest up to 40FPS/s.The post-processing method not only improves the accuracy of lane detection,but also reduces the false detection rate of lane lines.The accuracy of the method reaches 96.72% in the Tusimple data set.Meanwhile,the false detection rate and the missed detection rate are 2.89% and 2.60% respectively,which are significantly better than other models.There was also some improvement in the challenging Culane dataset.Finally,the validity of the proposed model is verified by the actual acquisition of the vehicle camera.