Uncertainty Informed Position Estimation Using Multi-view Cameras

Abstract Systems today rely heavily on sensor feed to inform an accurate operating picture. In the specific context of mobile robots and aerial assets, the integrity of this information has strong implications on risk assessment and safety. This work focuses specifically on uncertainty-aware data fusion applications in vision-driven tracking algorithms, with potential applications for various sensor modalities. We present a novel method for quantifying the component level uncertainty of parameters involved in solving the multi-view position recovery problem. Component-level uncertainty is quantified experimentally by comparing sensor data to ground truth measurements. Monte Carlo methods are used to simulate the propagation of uncertainty through the system, and predictions are presented in the form of a dense point cloud. The density and distribution of the resultant point clouds are used to approximate volumetric confidence intervals. System performance is evaluated and compared at different distance intervals, simulating expected performance as a system approaches a target.