Dynamic Object Tracking and 3D Surface Estimation using Gaussian Processes and Extended Kalman Filter

Dynamic object tracking is a prerequisite for the successful application of driving assistance and autonomous driving functions. With the advent of high resolution Light Detection and Ranging (LiDAR) sensors, more than one measurement emerges from the object's surface per measurement cycle which leads to an extended object tracking approach. The main research questions of this work are how to model the object's surface and the measurement model and how to estimate this surface recursively along with the kinematic state, e.g., position and velocities.We model the object extent in spherical coordinates as the radius over the azimuth and elevation angle in object centered coordinates. To this end we apply a recursive formulation of a Gaussian Process (GP) which estimates the surface radius and provides a very flexible and parameter-free model of the object surface. An augmented state space system is formulated with the state comprised of the kinematic state and the radius estimates for a set of basis points. Due to the non-linear measurement equation, an Extended Kalman Filter (EKF) is applied to recursively estimate mean and covariance of the state online over time. The proposed approach is evaluated in simulation given a point cloud model of an actual vehicle.Evaluation results show, that the GP formulation is capable of accurately representing any complex star-shaped surface in a parameter-free fashion. A recursive approximation of the GP enables recursive online tracking performance.