Reliable Estimation of On-Road Vehicles Motion Using Stereo Vision

Abstract

This paper proposes a 1st order extended Kalman filter for estimating the motion of on- road vehicles based on stereo vision observation of tracked point on the object's surface. The estimated states are the 2D position and orientation of an object relative to the ego-vehicle, the object's velocity, acceleration and the rotational velocity (yaw rate). The moving vehicle is reliably estimated in real-world dataset as in KITTI benchmark from within a moving ego vehicle. Practical on-road traffic situations such as oncoming traffic and turning vehicles at urban intersections were handled. The coordinate frames were assigned and the dynamic model and measurement model were formulated to be compatible with the estimation algorithms development and evaluation for the realistic KITTI dataset. Analytical effort was done for the characterization of the KITTI stereo vision measurement noise, while the dynamic models noise was solved effectively by the proposed mechanical limitations assumption. Practical issues such as filter initialization, numerical errors are covered fully. The overall system is systematically evaluated both using simulated and real-world KITTI data. The experimental results show that the proposed system is able to accurately estimate the object pose and motion parameters in a variety of challenging situations. The limits of the system are also carefully investigated.