Research On Road Detection Based On Deep Learning

With the increasing wealth of national material life and the improvement of per capita car ownership,frequent traffic accidents and road congestion have become the norm.In order to solve these problems,the development of intelligent systems such as automatic driving assistance system and vehicle navigation becomes more and more important.As one of the core technologies,road detection is paid more attention.Among many implementation technologies,machine vision technology has attracted much attention because of its relatively low equipment cost and simple equipment settings.Therefore,road detection method based on deep learning has become a research hotspot in recent years.Aiming at the urgent need of road detection,this paper studies the road detection method and road image generation based on deep learning.The specific research contents are as follows:1.Creation and collection of road database.A large number of urban road images and remote sensing road images are collected through vehicle camera,remote sensing satellite shooting,etc.the calibration software is used to calibrate them with the same group of students,and finally two data sets are formed: driving recorder dataset and remote sensing road dataset.2.A conditional random field based on convolution kernel is proposed.The theory of conditional random field and the popular fully connected conditional random field are studied,and the optimization algorithm of conditional random field is studied deeply.Aiming at the problem of error detection such as hole and edge unsmooth due to extreme light environment condition in road detection,a convolution kernel conditional random field is proposed,which combines the energy function of conditional random field with the full convolution regression neural network.Experiments are carried out on remote sensing road image and data set of dash cam.The experimental results show that the convolution kernel conditional random field combined with full convolution regression neural network can not only greatly improve the detection accuracy of the road,but also provide the optimization performance of random field beyond the full connection condition.3.The method of road image generation based on the generated countermeasure network.In order to solve the problem of insufficient road image in the data set of dash cam,based on the theory of generating countermeasure network,regression loss function with attenuation parameter and road detection loss function are added.The experiment shows that the improved generation countermeasure network can generate road image with higher quality.4.Research on road detection based on full convolution regression neural network.This paper analyzes some existing structures,such as convolution layer,expansion convolution layer,maximum pooling layer,average pooling layer,pyramid pooling structure,step-by-step connection and so on,to study the influence of these structures on the detection and prediction accuracy of two types of roads.Experiments show that for road detectionFor example,the use of skip connection can effectively improve the detection performance of the network;the network without pooling layer is likely to obtain the best detection performance;the expansion convolution compared with the standard convolution,the performance has been improved;the pyramid pooling structure has excellent detection performance,which can provide far better detection performance than the general network when the network can train parameters are similar.