Traffic Object Detection And Recognition Based On Deep Learning

Recent years,intelligent transport system(ITS)has gained extensive attention with the increasing of highway mileage and complexity of road networks.The detection and recognition of various traffic objects are key technologies of ITS,e.g.vehicles,road signs and licence plates.This thesis mainly studies the license plate recognition and traffic sign detection.Convolutional Neural Network(CNN)based methods are adopted in the vehicle detection and license plate recognition.In the stage of license location,the vehicle detection is conducted first,and then the license plate is located based on the detection results of vehicle detection.In this way,the background interference can be effectively suppressed and the license plates precisely locate.In the stage of license recognition,we improve the existing segmentation method of license plate characters based on the connected component analysis,and we use CNN for character recognition which improve the recognition accuracy.Our method can accurately locate and recognize the license plate even in complex scenes.The accuracy of license plate recognition system can reach 93%.We exploit Faster-RCNN to detect and recognize traffic signs.The tricks of divisional amplification detection,cross-layer connection and hard negative sample mining are employed to increase detection performance.Since the similarity in road traffic warning sign,we classify them with CNN based on the traffic sign detection results.The experiments on the laboratory data set,the average classification accuracy is 96%.The experiments on Tsinghua-Tencent 100K dataset show that the proposed method achieves 0.90 mAP,and the detection performance is comparable to other advanced algorithms.