| Perception system is one of the most important sub-systems of the Unmanned Automous Vehicles(UAV).In this paper,we focuse on the key technologies of road environment modelling including: lane detection,extraction of road boundaries and dynamical vehicles detection and tracking.All of the methods have been applied on the three UAVs of ourselves.The main contributions of this dissertation are as follows:1)This study proposes a real-time and robust lane detection method bases on the error prppagation model and Bresenham Line algorithms.First,to ensure the corresponding pixel of the lane with fixed width can be calculated precise,this paper introduces the perspective drawing method from Photogrammetry to construct a vanishing point based camera model for lane detection.Then,a high-confidence area for the image will be initilized to ensure the discrimination between lane markers and background noises.The error model of the perspective model is also taken into account.The features of lane can be figure out by the size information of the lane and a new Bresham Line Voting Space(BLVS)is used in this paper to calculate the characteristic line specific.Compared with the traditional Hough Transform(HT),the BLVS-based line detection algorithm can avoid a lot of floating operation which can improve the efficiency of algorithm.At last,an intremental regression method is used to estimate the geometrical model of the whole lane and Kalman Filter is used to tracking the parameters.2)Considering the temporal and spatial features of the 3D Light Detection and Ranging(LIDAR),this study introduces a region decomposition and cascading feature fusion method for road boundaries extraction.At first,the area of the road will be deconstructed by the density of the 3D LIDAR points.Cascading features are used to extract the drivable-region from coarse to precious and from near region to far region.Then the curbs of the road are located by resegmentation of the road regions and the BLVS based line segment detection method.With the line segments extracted from 3D LIDAR and images,the multi-lane model of the road will be created by a feature-level based fusion method.At last,in order to meet the real-time requirement of UAV,all the road boundaries extraction method will be calculated under the waiting time between two LIDAR bag.3)This study proposes a real-time obstacle detection and tracking algorithm based on a line segment detection method from point clouds of the vehicle contour and a deviation model to detection and tracking of the vehicles.At first,the line segment detection method of Edge Drawing(ED)for computer vision is introduced to extract the line segment of the contour of the vehicles in XOY plane.Then the geometrical model of the vehicles will be created by main line segment of the vehicle and the process of clustering will be efficiently operated.The line segments based vehicle clustering method is much more robust than the points based or grid based method.Then,the lane model in lane detection is introduced to construct the motion model of the vehicle.The departure distances to the lane boundaries are parameters of measurement of Kalman Filter and the position and direction of the vehicles will be accurate estimated. |
