Research On Ground Marking Recognition And Lane-level Positioning Using Multiple Sensors

Ground marking recognition and lane-level positioning is a very important research field in intelligent transportation systems.Ground markings consists of lane marking,ground traffic signs and so on.Drivers can replan or correct the path with the result of lane-level positioning.This paper did some researches on recognition of ground marking and lane-level positioning.Firstly,recognition methods on ground markings based on vision are carried out.First,calibration of camera is done to get the real-time inverse perspective mapping image.Perspective effect is then eliminated.Based on IPM image,recognition methods of lane marking and ground traffic signs are proposed.Then,to solve the problems of existed methods,such as high light,shadow,abrasion and obstacles,this paper proposed a lane marking recognition method based on custom vote space.Two-dim Gaussian filter and p-tile binaryzation are used for preprocessing in order to overcome problems mentioned above.Ground traffic signs' edge and sketch features are remarkable.Template matching and machine learning methods were used in existed researches.Hu moments and Zernike moments show low accuracy.By comparing template matching and machine learning methods,a multi-class SVM classifier using mixed Hu moments and Zernike Moments proposed in this paper won.To get lane-level positioning result,existed researches mainly use high-precision GPS or LIDAR while high-precision GPS is costly and LIDAR lacks of all weather performance.This paper proposed a low cost solution combing vision,millimeter-wave radar and low-precision GPS.A camera is used for lane recognition with circular curve model.A millimeter wave radar is used for road curb recognition by detecting the stationary railways alongside.Lane-level positioning is then calculated by comparing the relative distance of lane and road curb with the road prior information obtained from a low precision GPS.