High-precision Sementic Segmentation In Foggy Traffic Scene

Sense parsing is very important to automatic driving system in safety.At present,the main challenges of semantic segmentation in road scenes are the lack of data and the high requirement of robustness in different weather conditions and real-time in highspeed driving.In addition,the large scale difference of objects,the variety of objects,the changeable weather environment and the complexity in road scenes are also major challenges for semantic segmentation in road scenes.Therefore,using model improvement,data argumentation,data generation to improve the performance and generalization ability of the algorithm for road scenes,and to make the model more robust,is of great significance to the practical application of the automatic driving algorithm.In this thesis,classical algorithms in the field of semantics segmentation based on CNN are studied.And the structural advantages of position attention model are studied.More concretely,the position attention model can extract long-range dependencies information and improve object recognition accuracy.But there are also problems in position attention model,like information interference between similar categories and the over-focus on relevance information between objects which lead to the lack of detail information.For better semantic segmentation performance,this thesis proposes a hierarchical context information mechanism,which combines the long-range dependencies information and local context information(extracted from the hierarchical features)to enrich information and discriminate different types of semantic categories.In addition,due to the requirement of the efficiency of perception algorithm in road scenes,and the heavy computation burden of the position attention model,this thesis proposes an efficient position attention model.The proposed model reduces the scope of coverage of the model due to the complexity of road scenes,and deals with the problem that information extraction and fusion in complex scenes can easily introduce noise and affect the performance of semantic segmentation in road scenes.Finally,existing algorithms for semantic segmentation in bad weather use the finetune method which cannot make the full use of the relevance between defog task and semantic segmentation task.This thesis proposes a multi-task method,which combines the two tasks,and improves the security and robustness of semantic segmentation algorithm in foggy weather.