Collaborative Architectures for Relative Position Estimation of Ground Vehicles with UWB Ranging and Vehicle Dynamic Models

Abstract

The relative position to neighboring vehicles is critical to ongoing autonomy efforts including collision avoidance and path planning; therefore, it should not be fully dependent on an external reference such as GPS. This thesis presents methods for real-time relative positioning of ground vehicles by employing a network of on-board ultra-wideband (UWB) radios. The difficulties in range-based relative positioning, and the results from prior literature are described. Next, the proposed methods are derived which employ the kinematic bicycle model to constrain the estimated states to align with ground vehicle dynamics. The initial methods do not require vehicle-to-vehicle (V2V) communication. However, cooperative methods are also explored which make use of the simultaneous ranging and communication capabilities of UWBs. Feedback of the tracked vehicle’s dynamic states (velocity, yaw-rate, and steer angle) are analyzed for their impact on estimation quality. A geometrically-inspired consensus extended Kalman filter (CEKF) is also developed as a modification to both the prior work and the proposed vehicle-dynamic EKF (VehDynEKF). The methods developed in this thesis improve upon prior literature results in accuracy and robustness in the presence of UWB measurement errors, unfavorable relative geometry, and dynamic maneuvers. While the CEKF shows improvement over the prior literature methods without additional sensors, it under performs the VehDynEKF proposed here. With only the use of UWB ranging and odometry, the VehDynEKF in this thesis can provide robust relative pose estimates to a neighboring vehicle. The estimate is affected by relative dynamics but maintains a mean error less than 2.5 meters in both simulation and experimental results without cooperative feedback. The lateral velocity of the vehicles is found to be a primary contributor to error; odometry including measurements of the estimating vehicle’s lateral velocity significantly improves the results. Lastly, if the ego vehicle has access to the tracked vehicle’s longitudinal velocity, the mean error is refined to be less than 1 meter—sufficient for the majority of safety-critical applications.