Extraction Of High-precision Line Object On The Road Using Mobile Mapping

The autonomous vehicles(AVs)have been developed rapidly due to the national strategy.High-definition(HD)maps,as the core technology of AVs,attract more and more attention from markets and research institutes.Road and lane marking,which provide driving area for AVs and guide them to drive safely,are important elements of HD maps.As the data collection device of HD maps,mobile mapping system(MMS)commonly integrates different types of sensors,such as camera,laser rangefinder(LRF),inertial measurement unit(IMU),and global navigation satellite system(GNSS),and provides different types of data for HD maps.It is an important research field of HD maps to extract high-precision roads and lane marking with the combination of different types of MMS data.However,the complex and diverse types of the road boundaries,the interference information on the road,and the wear,blurry and missing lane markings make the extraction of the road and lane marking high challenging.In this thesis,using different sensors of MMS and the data captured by them,extrinsic calibration between sensors and the high-precision extraction of the road and lane marking are studied.(1)The extrinsic calibration between a camera and a 2D LRF and the extrinsic calibration between a camera and an IMU are studied.The laser range data captured by the 2D LRF record only one scanning line,which is formed by the intersection points of the laser scanning plane and the object surface.The image obtained by the camera is a 2D imaging of the 3D object.In addition,the laser range data is invisible in the image since the laser used by the LRF is outside the visible spectrum.Therefore,there is no corresponding feature in these two types of data,and the extrinsic calibration between the camera and the LRF cannot be implemented by establishing an accurate correspondence between them.In this paper,a photogrammetric control field is used to obtain the corresponding points between the image and the control field and those between the laser range data and the control field,respectively.By establishing the accurate correspondence relationship with corresponding points,the accurate and robust calibration of the camera and the LRF with respect to the control field are implemented respectively.Thus,the accurate and robust calibration results between the camera and the LRF are obtained.Then,for the calibration accuracy affected by image blur,the extrinsic calibration method between the IMU and the camera is improved based on the kalibr,a calibration toolbox for IMU and camera.The experimental results demonstrate our method can improve the accuracy of the calibration between the IMU and the camera.(2)Based on Gestalt Perception Theory,the adaptive extraction method of the road is studied.Road extraction is an important part for HD maps production procedure.However,roads have complex and diverse boundary types such as vegetation,soil and so on.Therefore,adaptive extraction of roads with different types of boundaries is a challenging and urgent problem to be solved.Point clouds captured by the MMS are unstructured and unordered.The road can be perceived in point clouds because the road point clouds are regular and form the most significant plane.According to the Gestalt Perception Theory,the more significant the shape is,the smaller the probability of the shape appearing is.Inspired by this theory,the road extraction in point clouds is modeled as a probability problem exploiting the plane significance of the road in this paper,and roads with different types of boundaries are adaptively extracted by finding the minimum probability.(3)An extraction algorithm of lane boundaries combining panoramic image and point clouds is studied.Extracting lane marking with high precision is an important research field of HD maps.However,the interference factors,such as interference lines,blurry lane markings and missing lane markings,increase the difficulty of lane marking extraction.The image has rich textures,so the lane marking and the road in the image have obvious differences,but it cannot reflect the 3D geometric and location information of lane markings.Additionally,the distortion caused by the projection in the image also matters.Although point clouds compensate the disadvantages of 2D images,the lack of texture information in point clouds makes the blurry lane marking almost invisible,especially in point clouds without reflection intensity,the position of the lane marking cannot be recognized.With panoramic images and point clouds without reflection intensity information collected by MMS,this paper proposes an extraction algorithm of lane boundaries combining the texture feature of panoramic image and the geometric feature of point clouds.By analyzing and utilizing the edge features of lane boundaries in panoramic images and the geometric constraints they have in point clouds,the normal and blurry lane boundaries in the complex road environment are extracted.At the same time,the missing lane boundaries are supplemented.