De-noising Motion Predictions of Scuba Divers for Aquatic Robots

Abstract

Current diver predictors output a sequence of bounding boxes, with two corners randomly sampled from two bivariate gaussians. This introduces noise and uncertainty into the prediction outputs and makes the conversion of this sequence of 2D boxes into a 3D motion vector challenging. This poster describes an approach to de-noise this output and convert the predictions into a format that can be used by aquatic robots to plan their motion and follow scuba divers robustly.