Uncertainty Estimation for Stereo Visual Odometry

Abstract

Over the past few decades, unmanned aerial vehicles (UAVs) have been increasingly popular for use in locations that are lacking, or have unreliable global navigation satellite system (GNSS) availability. One of the more popular localization techniques for quadrotors is the use of visual odometry (VO) through monocular, RGB-D, or stereo cameras. With primary applications in the context of Simultaneous Localization And Mapping (SLAM) and indoor navigation, VO is largely used in combination with other sensors through Bayesian filters, namely Extended Kalman Filter (EKF) or Particle Filter. This work investigates the accuracy of two standard covariance estimation techniques for a feature-based stereo visual odometry algorithm. An effort is then made to learn the odometry errors by means of Gaussian process regression (GPR). By evaluating positioning error while monitoring VO confidence metrics (e.g., amount/quality of features tracked, motion between frames), an estimate of the uncertainty in the VO position estimate is realized. The experiments carried out in this work are first performed in a ROS/Gazebo simulation environment, where the true position of the UAV is known and can be compared directly against the VO estimated position allowing for meaningful conclusions of the covariance estimate. This is valuable information for filtering strategies and motion planning under uncertainty algorithms, especially when the environment is not consistently rich in features. Proper knowledge of the covariance in the estimate can lead to neglecting the motion terms with high position uncertainty, preventing the VO solution from negatively effecting the estimate of the filter. The experiments in simulation are then extended to experimental testing on the UAV hardware setup with a Vicon motion capture system as ground truth.