Key Technology Of Semantic Segmentation Of Crossing Weather Road Scene

With the development of research on driver assistance systems and driverless cars,vehicles are becoming more and more intelligent.Smart vehicles initially relied on sensors to collect environmental data around the car,which in turn controls the car's response to the surrounding environment.However,these sensing devices are not only bulky but also expensive.In this case,the vehicle needs to be able to "understand" the road scene as the driver can.Therefore,the demand for vision based road scene understanding is more and more urgent.In conclusion,it is necessary to study the key technologies of semantic segmentation of road scenes to promote the practical application of autonomous driving system.Presently,road scene segmentation in autonomous vehicles has achieved good results in good weather and normal lighting conditions.However,the semantic segmentation effect of road scenes is not very good in the case of poor weather and inadequate lighting and low visibility.How to realize semantic segmentation of images under complex road conditions is an urgent problem to be solved in the development of intelligent vehicles and is the focus of this paper.The research of this paper is divided into two parts:image retrieval and semantic segmentation.The main contents are:(1)Aiming at the problem of unsatisfactory semantic segmentation in road scenes under complex and changing weather and light conditions,this paper proposes a method of image retrieval and semantic segmentation across weather scenes.Road scene images from different data sources are collected and marked Includes two driving directions and scene photos taken from AMOS.(2)A Metric Learning-based approach is proposed to construct a deep convolutional neural network structure of SCNN + ResNet to extract the image features and minimize the impact of weather and light on the features,so as to obtain the ability to represent the same location under different weather and light conditions of the road scene image features,in order to facilitate the image retrieval.On the basis of SIFT-Flow,this paper proposes MRF to fuse label information,priori spatial information and smoothing information to establish the dense matching relationship of pixels so that the semantics can be transferred from the known image to the target image to obtain the semantic segmentation result of the target image.(3)Experiments are carried out according to the proposed method,and the results of image retrieval and semantic segmentation are analyzed and evaluated.The experimental results show that the accuracy of image retrieval is greatly improved under the high contrast condition,and the key information points(such as traffic signs,traffic lights,etc.)of the proportioned road scenes occupying a small area of the image are exhibited a good segmentation results.