Traffic Sign Recognition Research Based On Deep Attribute Learning

Traffic sign recognition plays an important role in intelligent transportation and autonomous vehicles.However,general and accurate traffic sign detection is still a challenging task due to the influence of light and complex background.The existing traffic sign detection methods have achieved good results,but there are still some shortcomings.On the one hand,when detecting candidate regions of traffic signs,the existing methods do not take into account common characteristics of different colors on the surface and the materials of traffic signs.On the other hand,the prior knowledge will be used to filter non traffic sign regions after candidate regions are classified.Non traffic signs include traffic lights and signs along the roads.The existing candidate region extracting methods are not universal and traffic sign recognition methods are not robust enough to deformation and complex background.To overcome these limitations,we propose a novel deep attribute learning based traffic sign recognition method.In detection of traffic sign candidates,we fully explore the common characteristics of different colors on the surface of traffic signs and the reflectivity of materials they are made by.Based on these characteristics,traffic sign regions are connected areas respectively in the three gray images generated by the three components of HSV color model.Therefore,we detect connected areas to extract candidates of traffic signs in the three gray images respectively.Firstly,image preprocess is operated.The image is converted to HSV color space,and then three channels of HSV are used to generate three gray images respectively.Secondly,the connection domains are extracted as the traffic sign candidate regions on the three gray images respectively.This method is not only specific to a particular color(such as red,green),but also use no prior knowledge like the shape of traffic signs.Hence,this method is simpler and more universal.In recognition of traffic sign candidates,three attributes are introduced:shape,color and patterns of traffic sign.We add attribute learning constrains to convolutional neural network learning so that attribute learning and classification learning are carried out simultaneously.The three attributes describe semantic information of traffic signs and bridge the gap between low level pixels and top level semantic information of images,which makes machine easier to understand traffic sign images and to predict the category of traffic sign more precisely.Since more than one candidate regions are detected,which cover the same traffic sign,during traffic sign detection,traffic sign region filter is used to remove the redundant regions.Experiment results on Sweden Traffic Sign Detection Dataset(STSD)and German Traffic Sign Detection Dataset(GTSD)show that the proposed method can effectively improve precision and recall in terms of traffic sign recognition.