Traffic Sign Detection In Video Using Deep Neural Network

Road traffic signs are indicative signs to ensure driving safety and traffic smoothness.Every driver on road should observe the instructions of traffic signs.However,traffic violations and even terrible traffic accidents,which is endangering safety of people and property,occur because of ignoring or misjudging traffic signs.Therefore,the research of traffic sign detection algorithm is particularly important.With the development of computer vision theory,convolutional neural network is used in the field of objection detection,especially in the subject of traffic sign detection.Traffic sign detection algorithm which is applied in unmanned driving or assisted driving has become a hot research topic at home and abroad.Besides,it is also a practical need to improve driving safety.However,with the required accuracy of objection detection products is increasing,the detection through single frame can not meet the actual needs.Therefore,the method using continuous frames in video for traffic sign detection become a new solution.There are three main points in this thesis.Firstly,aiming at the problem that there is no proper data set for experiment,we sorted out the existing data sets and relabeled them.Moreover,we collected videos taken on the actual road and annotated them.Secondly,feature pyramid network is used to fuse shallow features with deep features.Detection on multi-scale feature map optimizes the detection performance of detection algorithm for small-scale traffic signs.Thirdly,aiming at the problem that the accuracy of traffic signs classification is low,we proposed an improved algorithm,which combines the feature information of continuous frame images to strengthen the classification.Finally,the algorithm proposed in this thesis can achieve a higher accuracy of traffic sign detection,which can be applied to the actual driving environment.It has practical value.In order to analyze the performance of the proposed algorithm,we designed several comparative experiments.They demonstrated the advantages of the feature pyramid network and the improved network get an improvement of 5.6% in accuracy.