Multiple Features Cooperation For Traffic Sign Detection And Recognition

Traffic signs are the facilities which transform specific traffic information to observer by using color,shape,characters or graphics.It can help to regulate the order of the traffic.Automatic detection and recognition of traffic signs is one of the most important parts in intelligent driving assistance systems and unmanned vehicles.So the research has practical significance and enormous application requirements.Accurately detection and recognition of traffic sign were still a challenging topic in complex traffic scenarios.This thesis is mainly focus on developing the recall and precision rates of detection in video which captured from moving vehicles.The main work of this thesis include two parts:A new method was presented to improve traffic sign detection in single traffic image with cooperation of color,shape and scale features,especially under conditions of color distortion,shape deformation and scale variance.Color enhancement maps were generated from traffic scene images.Regions of interest were then extracted from the color enhancement maps using multiple thresholds of color,chain codes of the curvature histograms of closed contours are calculated and scale normalized for the contours.The SVM(Support Vector Machine)classifier was applied to classify the chain codes of the extracted traffic signs and the template signs.Our experiment results demonstrate that this method is capable of improving recalling rates under normal and the three specific conditions,with low time complexity.Based on the results obtained from the detection algorithm mentioned above,a new method which fusion the depth information of scene was proposed to detect traffic signs in stereo video sequence scene.The 2D appearance of the same traffic sign in adjacent frames are associated based on the projection consistency with depth information.Then the stable relationship between the projection size of traffic sign with the local depth information was used to filter the unreliable candidates in pursuing higher precision rates as well as recall rates.Experiments were carried out based on a public video dataset captured by a multi-views camera system.The Statistical results indicated that in guarantee high recall rates,the precision was significantly improved in fusion depth information.The final recognition results were determined by the highest score of a feature matching scheme in between bag of SURFS with templates'.