Lane Line Detection And Recognition Based On Deep Learning

The automatic driving task of driving vehicles can effectively solve the problems such as traffic.It is a very hot and important research topic at present,and lane line detection and recognition is one of the key technologies.Using deep learning technology instead of traditional image processing technology,a lane line detection and recognition model with high recognition rate,strong robustness and high timeliness is realized.The research contents and methods are summarized as follows:(1)Study the current mainstream lane line recognition model,including the implementation steps based on the traditional image processing model and deep learning model,and give a graphical representation of the model structure.(2)In order to extract the lane line feature information better,the ARB neural network module is proposed for the lane line detection and recognition task.(3)Study the Deeplab v3+ network structure,embed the ARB module on this basis,and design the end-to-end lane line detection and recognition network.(4)Use the F-Measue value to evaluate the recognition of the lane line and compare it with other deep learning models.(5)Combining clustering and inter-frame correlation technology,a fast calibration algorithm for lane line position is proposed to realize fast calibration of lane line position.(6)From the three core parts of the ARB module's different combination structure,window sliding efficiency and clustering efficiency,analyze the timeliness of the whole system.The experimental results show that the designed ARB network module can effectively extract the lane line feature information in the strong illumination,dim light,shadow occlusion,traffic congestion road,road damage,lane line missing,curved lane line and normal road environment.Experiment on the CULane public dataset,calculate the F-Measure value to 0.9731,which is 2% ~ 5% higher than other deep learning models.The proposed lane line position calibration algorithm controls the lane line positioning efficiency to 40 msec/frame.Within,it is nearly 20 milliseconds/frame compared to violent search.The proposed ARB network module breaks through the convolution in the traditional single direction,convolves and merges from different dimensions of the feature map,and has better feature extraction effect on the special structure of the lane line.The experimental results of the ARB network module show that for the recognition target with special structure,the convolution method can be changed,the target feature information can be analyzed from multiple dimensions,and the redundancy dimension can be further simplified through comparison experiments.This targeted network design structure can express the feature information of the target more comprehensively and efficiently.