Research On Traffic Sign Recognition&State Tracking And Estimation Algorithms In Complex Environments

Since the beginning of the21st century, many researchers and manufacturers invested a lot of human and material resources in the research and development of unmanned vehicle’s technology. Road condition perception based on computer vision is one of the important researched contents:recognition on the information of traffic sign and the state of traffic light in both sides of the road. It will provide the decision-making basis for driving the unmanned vehicle.On the basis of existing research results in the detection, recognition and tracking of the traffic signs at home and abroad, combining the practical application and test performance of the traffic recognition system in the unmanned vehicle, the real-time system of traffic sign recognition and tracking for the unmanned vehicle was designed and constructed. The main research work and achievements are as following:An imaging process of camera was introduced here. According to the original parameters of CCD(Charge Coupled Device) and intrinsic parameters by the camera calibration, the approximate distance and angle between the traffic sign, and the position of camera on the vehicle are estimated in the road environment. In order to control the exposure time of the camera and adjust the brightness of the image, the brightness information and the exposure value of region which the traffic sign appeared are judged in advance. Different weight matrice are chosen by the exposure of image to calculate the best exposure time. The proposed method can accurately adjust the exposure time of the camera and get image which has appropriate brightness. It is suitable for detection and recognition of the traffic sign.The detection and recognition algorithm for50symbolic and text traffic signs are presented. RGB value of the image is transformed to highlight the characteristic colors(red, yellow, blue) of the traffic signs. The appropriate thresholds are selected for the image segmentation. The edge of traffic sign’s ROI(region of interest) is reconstructed to decrease the error rate. The shapes of traffic sign’s ROI are classified and excluded the interference by the signature. For text traffic sign, atrovirens and blue regions are segmented by the thresholds. The regional morphology is judged, and the rectangle region is selected as a traffic sign candidate. The binary image of the candidate region projects to horizontal and vertical direction. The projection curve is fitted by cubic spline interpolation; the curve peak is located to determine the positions of the row and column to separates the text region. Two model representation methods:(1)two dual-tree complex wavelet transform and two-dimensional independent component analysis;(2)template matching based on the internal graphics recognize the traffic sign’s type of candidate region respectively. Then the two recognition results are fused and excluded by decision rules. The experimental results show that the recognition rate of traffic sign is more than91%, and the average processing time is171ms. The proposed algorithm has superior performance and is suitable for the sensing system on unmanned vehicle.Traffic light detection and recognition algorithms for arrow and combination type are proposed. For the arrow traffic light, the position of the board and lamp of traffic lights are located by the color and morphological features of the board and lamp. For the combination of traffic light, the image is transformed by TopHat and convered to YCbCr from RGB space firstly. Then it is segmented by the threshold and filtered by morphology. The regions are preliminarily filtered according to the morphological information such as width, height, area, and duty radio. The ROIs are combined and projected to the horizontal direction. The single region of traffic light is located by the valley. In the traffic light recognition process, the candidate region image is grayscaled and normalized. Feature is extracted by the Gabor wavelet and reduced the redundancy by2dimensional independent component analysis. Feature is sent into the nearest neighbor classifier to judge direction information of the traffic light. Many videoes are collected in3cities and tested algorithm performance. Comprehensive performances show that overall recognition rate of the algorithm is more than91%and the average processing time is152ms. It achieves real-time, stable, accurate goal to recognize traffic light.The multi-target tracking model of the traffic sign is established. And the target state of the traffic sign is defined. The target state and trajectory model of one traffic sign is established by using unscented Kalman filter to predict the position of traffic sign target. For the traffic light, the board and lamp of the traffic light are tracking by the Kalman filter. Observation sequences are selected to train Hidden Markov Model parameters for one and three traffic light. The next state is estimated by the hidden Markov model algorithm.The experimental platform of traffic sign recognition system is established. Two subsystems of traffic sign recognition are designed and implemented:(1) the traffic sign recognition system;(2) the traffic light recognition system. It detailedly performs the system function, display form and parameter setting.