Research And Implementation Of Lane Detection Based On Deep Convolutional Neural Network

Lane detection is one of the most important components of driving assistance system.With the increasing demand for cars,traffic accidents occur frequently,among which more than 50%are caused by illegal lane changes.Lane detection technology has been widely concerned.Accurate and fast lane detection technology is of great significance for lane line departure alarm,lane line maintenance and other functions.Lane detection technology needs to be accurate and real-time.Traditional lane detection methods are too dependent on manual design features and a series of heuristic post-processing techniques to be directly applied to complex road scenes.In recent years,deep learning technology has performed well in the field of computer vision,and some scholars have begun to apply it to lane detection tasks,and achieved better results than traditional methods.However,the current lane detection technology based on deep learning still has some flaws.For example,the lane lines are easy to be interfered by the covering or affected by the shadow and light,so it is impossible to train a better model just by the features of local pixels.In addition,deep learning needs to train a large number of parameters,which will not only affect the operation speed of the model,resulting in the detection time is too long to carry out real-time detection,but also cause the model is too large to be deployed in terminal equipment.In view of the above problems,this paper uses the method of deep learning to study the lane detection technology,improves on the basis of the existing network,improves the detection speed of the model,and compresses the model.The main work is as follows:(1)Firstly,this paper summarizes the current mainstream lane detection methods,introduces the relevant theoretical basis and makes a comparison.After analysis,this paper builds the model with reference to Mask-RCNN.(2)In this paper,ResNet-50 is used as the convolutional backbone of Mask-RCNN to detect lane lines,and ResNet-101 is used as the convolutional backbone for experimental analysis.(3)The shallow neural network is designed as the convolutional backbone of Mask-RCNN,and it is optimized to reduce the training parameters in the network,so as to achieve the goal of high-speed processing.(4)In this paper,the model is compressed by the quantization of fully-connection layer and convolutional layer,so that the model can run on terminal equipment.