A computational geometric learning approach for person axial and slanting depth prediction using single RGB camera

The paper proposes a non triangulation method for estimating the depth of a person in slanting position relative to camera lens center. The influence of three parameters, namely lens aperture radius, focal length, and object size on person depth along the axial line, is well investigated while recording the in-focused portions of the person at various distances on the axial line. This investigation is taken as a basis for estimating the depth of the person on the axial line for different combination of aforementioned parameters using machine learning framework. Further to estimate the slanting depth, a geometrical relation is obtained from the predicted axial depth, deviation taken by the person to the left/right of the axial line. Considering the ground truth depth data taken within 3 to 6mts range using Nikon D5300, the model is validated and is inferred that the camera to object distance (or depth) anticipated along axial line is 99.06% correlated with actual camera to object distance at a confidence level of 95% with RMSE of 17.36