Multi-dimensional Road Modeling And Vehicle Localization Optimization In Vision-based Driver Assistance System

Vision-based driver assistance system (VB-DAS) uses camera as the main sensor,and aims for safe driving of vehicles. According to the research objects, DAS canbe divided into two general directions: in-vehicle driver monitoring and out-vehicleenvironment perception. In-vehicle driver monitoring mainly uses cameras to studythe face images of the driver, and monitor his/her mental status, including fatigue stateand gaze direction, to assure efficient control of the vehicle when in driving. Whileout-vehicle environment perception tries to sense and analyze the objects around thevehicles, such as roads, pedestrians, other vehicles and so on. By doing so, blindspots of the driver will be eliminated, and hidden danger can be avoided. As roadis the main environment for a vehicle, road detection and modeling is a hot researchtopic in DAS. Moreover, information of ego-motion is of significant for a vehicle, asit can help to assist vehicle to understand motions of its surrounding objects. Thepaper studies multi-dimensional road modeling and detection in vision-based DAS, aswell as its quantitative accuracy index; discusses the fusion of vehicle’s ego-motioninformation to road modeling, and its effect to improve the accuracy and adaptationof road modeling; put-forwards a model for vehicle localization optimization basedon dynamic tracking of vertical road profile, and proves the effect of utilizing roadmodeling to suppresses the drift of vehicle ego-motion estimation.As a common road perception method, lane detection is first discussed in the pa-per. Widely used lane models, assuming a global geometry shape of lane boundaries,will be insufficient in detecting lanes with various geometries in real road situations.A new three dimensional parameterized free-shape lane model, fully utilize the threedimensional constraints between two-boundary lane’s border, is put forwarded in thepaper. With the new lane model, lane detection and tracking method based on point- by-point detection is also suggested. Experiments using different road conditions havevalidated the adaptation of the new lane model to the various geometry shapes of lane,as well as the sufficiency and robustness of the detection and tracking method.The concept of “lane” will be insufficient on road conditions when there are mul-tiple or no lane boundaries. A concept of “corridor” is put forwarded in the paperas a new road perception method to describe road surface with any condition of lanemarks. By fusing vehicle’s ego-motion information with lane marks,“Corridor” de-fines a corridor-like road area with constant width in the lookahead distance. The areafollows “vehicle drives in lane” assumption, and will be driven through safely by thevehicle. Following corridor concept, corridor detection and tracking method is alsoprovided in the paper. Experiments on various road conditions prove the benefits ofthe corridor concept, as well as the sufficiency and robustness of the corridor detectionmethod. Moreover, it is shown that vehicle ego-motion is of usefulness to improve theadaptation of road modeling methods to road conditions.When vehicle is driving on an outdoor road, vertical road profile modeling willbecome necessary, when the road is downhill, uphill or uneven, as the commonly usedplanar road assumption will be failed. Compared with linear line or parabola, b-splinecurve is more powerful to model the vertical road profile, and suits the situations whenroad is up-or down-hill, or have complex height variation. The paper adopts Siftflow to obtain the3D information of feature points on road surface, and uses b-splinefor curve fitting of the feature points. Experimental results show that the suggested b-spline vertical road modeling is effective to describe various vertical road conditions.In order to test the accuracy of vertical road modeling result, an inverse perspectivemapping for general road with vertical road model is suggested. Parallelism of theparallel lane in bird’s-eye view is used as accuracy index for vertical road profile mod-eling.Localization drift from relative error accumulation is a common phenomenonis visual odometry. A new drift distribution model is suggested in the paper. Thenew drift model establishes the relation between propagated error and the trajectorytraveled in ego-motion estimation, and provides the drift distribution for every pointin the trajectory. The drift model provides a general error analysis method for visual odometry, and adopts for drift performance comparison of various visual odometryalgorithms. Three different visual odometry algorithms are conducted in monte carlosimulation, and the results prove that the theoretical drift distribution model developedis consistent with results from Monte Carlo simulation.Finally, the paper put forwards a model for vehicle localization optimization fromvisual odometry, based on dynamic tracking of vertical road profile. As there aresome lags between road modeling and ego-motion estimation, a fixed-lag kalman filterframework is adopted. Based on the assumption that a vehicle will always on the roadsurface in driving, the state update model and measurement model for the kalman filterare discussed. Results show that the drift in navigation can be effectively suppressedby result from vertical road profile modeling.