Deep Learning Based Scene Recognition For Autonomous Driving

With the development of social economy,due to traffic infrastructure construction and road conditions can not meet the rapid expansion of traffic demand,road traffic safety issues and traffic congestion problem have become stumbling block to improve the efficiency of social production activities.Unmanned vehicles have a broader environmental perception than humans and unique ability of route planning,which can reduce the traffic accident rate and ease the traffic congestion pressure.It can be seen that the development of unmanned vehicles has great practical significance for alleviating and even solving traffic safety problems and traffic congestion problems.Scene recognition for autonomous driving refers to the use of car camera records of vehicles around the visual environment data,identify the current autonomous driving traffic scene and environmental information,it is one of the important tasks of the unmanned vehicle system sensing layer.As the unmanned system relies on environmental perception results for driving behavior decisions,the output of scene recognition will have a profound and dramatic impact on the driving behavior of unmanned vehicles.Therefore,the study of autonomous driving scene recognition is of great practical significance.In recent years,with the rapid development of the deep learning,solutions based on the deep learning are widely used in a variety of scientific fields and achieved good results.In particular,the Convolutional Neural Network(CNN),because of its sparse connectivity and shared weights characteristics,reducing the number of weights while reducing the complexity of the network model,it is particularly suitable for the task with large amount of input data and rich information,such as image scene recognition.For deep larning,datasets with density and diversity is the key factor in achiveving good scene recognition result.But there is no large-scale database for autonomous driving scene recognition,which seriously restricts the development of deep learning in the field of autonomous driving scene recognition.Based on autonomous driving typical scene category,this paper builds the world's first large-scale dataset for autonomous driving scene recognition,named “SYSUDrvingScene(SYSU-DS)”.The dataset contains 282700 marked images of autonomous driving scenes,consist of 4 subsets,57 scene categories,more than 950 thousand scene labels.SYSU-DS is the largest and most complete range of autonomous driving scene recognition datasets,providing a solid data base for autonomous driving scene recognition.In this paper,we test and publish a variety of advanced deep learning classification model through the SYSU-DS dataset,verified the existence of a reasonable deviation and certain degree of generalization of the dataset.SYSU-DS dataset has both professionalism for autonomous driving scene recognition that able to objectively and truly reflect the real scene information,and a reasonable scene recognition accuracy rate for models trained on other dataset.In this paper,following the center loss idea to formulate the similarity aware loss,we introduce a scene similarity aware deep neural network,named SceneNet.By comparing with other state-of-the-art scene recognition networks,Scene Net got better autonomous driving scene recognition results,solved the problem caused by autonomous driving scene which has both the similarity of different scenes and the difference in the same scene.In this paper,scene-baesd image retrieval experiment and scene image multi-label tagging experiment were carried out through the SYSU-DS dataset.Seceral state-of-theart image retrieval models and multi-label tagging models yielded reasonable experiment result.It is proved that the SYSU-DS datasets has the feasibility of research topics besides scene recognition,such as scene-baesd image retrieval and scene image multilabel tagging,which proves that the SYSU-DS datasets has great academic value and application potential.