Pose estimator for autonomous vehicles

In autonomous vehicles the state identification of the vehicle itself is a key task, through it position and orientation of the vehicle in the working environment can be inferred, and control logic can perform the needed control action on these information. Therefore it is clear that the quality of the state estimation affects directly effectiveness of the whole control. This thesis deals with the pose estimation of a small-scale autonomous vehicle demonstrator, also addressing the real-time implementation of the algorithms on an embedded platform. The system is designed to work in a demo environment that can be either indoor or outdoor, hence this application kind leads to a state formulation based on a curvilinear abscissa reference framework, namely composed by the advancement along the path, the yaw and the lateral distance w.r.t. the path; which is assumed to be the lane center. The first quantity can be estimated through odometry, while the last two can be estimated through vision. This work focuses on these latters. To get the lateral position an the relative yaw, as first step, a lane detection system has been developed. To fulfil the real-time requirement, identification of the lane’s boundaries is performed through a tracking algorithm, and all software has been written in C++ language (with OpenCV library). From the detected lane boundaries the lane center is first estimated on frame, and then projected into real world. Then, fitting a clothoidal model on lane profile, vehicle’s offset and yaw w.r.t to the path are estimated. Real time feasibility and reliability of the algorithms have been proven off-line on a simulation video running on an O-DROID XU4 board.