Research On Road Detection Method Based On Deep Learning

Today,road detection is still a challenging task in intelligent driving.For an image containing road areas,obstacles,and pedestrians,the primary goal of a road detection task is to accurately detect road areas while avoiding obstacles and pedestrians.This determines whether autonomous driving technology can bring enough safety to people.With the continuous development of computer vision,more and more deep learning methods are applied to road detection tasks because they can obtain deeper picture features and discover roads from the original RGB pictures.However,due to the effects of lighting and background environment,the road regions extracted by these methods are still not accurate enough.Therefore,the accuracy of road detection algorithm needs to be improved.At present,the more popular road detection methods include traditional methods based on sensors and deep learning methods based on neural networks.In general,detection methods using deep learning are more accurate and robust than traditional methods.This paper analyzes these methods and proposes a road detection method based on deep learning.The main research contents and innovations of this article are as follows:(1)Introduce the cycle-consistent adversarial network to complete the road detection task.The network can find two mapping relationships between one image set and another image set,and these two mappings can complete the feature conversion of the image.It has been widely used in defogging and snow removal.This paper uses this network to find the mapping relationship between images of undetected roads and images of detected roads,and completes the task of road detection after feature conversion.(2)Aiming at the common problems of false detection and missed detection in road detection,this paper proposes a new model.The main reason for missed detection is light.Too bright or too dark environment will make the characteristics of missed detection area and road area too different.The false detection is mainly because the characteristics of some walls or obstacles are very similar to those of roads.Therefore,the model adds connected domain detection and flood filling,which makes up for missed detection areas while excluding false detection areas,thereby solving most of the problems of false detection and missed detection,and improving the accuracy of the method.This article selected the KITTI dataset during the training and testing phases and uploaded the test results to the KITTI official website for further evaluation.By analyzing the returned evaluation data,we can conclude that our method has a high precision and a good robustness.