A supervised learning approach to far range depth estimation using a consumer-grade RGB-D camera

This paper describes an approach to learn far range depth data from a consumer grade RGB-D sensor. Raw sensor depth data is limited to the near field and is used as ground truth by a supervised learning algorithm for predicting far range depth. A multiple hypothesis regression function is fit using the near range ground truth. A linear least squares solution determines the parameters of each function in the hypothesis vector. The “best fit” function is chosen as the one which minimizes the error between the predicted depth and the actual depth. This is used to populate a lookup table with learned far range depth data, extending sensor range. A novelty detection scheme estimates change in terrain inclination and is used to repopulate the lookup table of depth data for varying ground plane inclinations. Learning depth extension has been tested in heterogeneous indoor and outdoor environments, with diverse terrain type, varying terrain inclination and height, and under different obstacle field configurations.