| As a strict established road facility,traffic signs play an important role in reflecting local road conditions,evacuating traffic and ensuring safe and orderly driving environment.With the improvement of modern intelligent driving technology,the automatic detection and recognition of traffic signs has become a hot topic in the field of driverless.Traditional machine learning algorithms can only capture a single feature of traffic signs.Moreover,they mostly rely on prior processing and experience judgment,thus not universal.Instead,The method based on deep learning can mine more layers and more abundant semantic information of traffic signs by constructing Deep Convolution Neural Network(DCNN)model,so as to realize intelligent detection and recognition and meet the requirements of real-time and accuracy.In this paper,based on the full study of the advantages and disadvantages of one-stage and two-stage target detection network,as well as a variety of multi-scale detection methods,two deep learning methods for traffic sign detection and recognition are proposed.One way is to construct a residual SSD(Single Shot Multibox Detector)network,and the coarse-to-fine multi-scale blocking strategy is adopted to realize the detection and recognition of traffic signs.Firstly,multi-scale blocks for high resolution traffic images are presented.Then the middle-scale image blocks are roughly detected,and afterwards the other scale image blocks related to the coarse detection results are put into the network for fine detection.Hard negative mining is used to keep the balance between positive and negative samples in training model,and the prediction results are optimized by twice non-maximum suppression separately at the level of original image and image block in testing.Experiments show that the coarse-to-fine method has high detection accuracy,but the whole process takes a long time because of the large number of multi-scale image blocks.Another way is to construct a cascaded network to realize the detection and recognition of traffic signs.Firstly,a single scale block is used to reduce the number of image blocks,and then a cascaded network structure,cascading a two-class network after RetinaNet network to reduce false positive instances,is adopted.The FPN structure of RetinaNet network fuses multi-layer feature map information,and the two-class network can make supplementary discrimination of foreground and background for the first stage prediction results,so the prediction accuracy isimproved greatly.Focal Loss classification function is used to solve the class imbalance problem in training model,and some criteria are used to synthesize the prediction results of two stages of cascaded network in testing.Experiments show that this cascaded network method achieves a good compromise between accuracy and time-consuming,and can achieve the basic balance of different classes of prediction. |
