Depth Perception in a Single RGB Camera Using Body Dimensions and Centroid Property

摘要

Infrastructure Supervision is a compelling need for buildings and open areas. It is facilitated through the joint use of stereo vision cameras, techniques and algorithms. This Stereoscopic assessment helps monitoring systems to reconstruct people's visible surface and also provides a robust estimation of the position and posture of the person that allows 3D scene activities and interactions. In practice, in occluded fields, the correspondence between pixels and pixels interferes with the flow of data in surveillance. Structured light ToF imaging and Light Field imaging sensors came into being considering the restriction. These techniques, however failed in addressing the inaccuracies and noise introduced in the phase of profound capture. Based on the Human Anthropometric research, we suggested a technique for estimating the depth of an individual from a single RGB camera. As we deal with moving objects in a scene, also consideration is given to centroid ownership. The system is trained by feeding stature, body width and centroid as inputs to estimate a person's actual height using gradient boosting model. And a person's further anticipated height and actual height are used to predict distance. After taking actual depth (camera to person distance) and real height as ground truth, the suggested model is validated and it is inferred that the camera to person distance anticipated (Pred(dist)) from estimated Real height is 95% correlated with actual Camera to Person distance (or depth) at a confidence level of 99.9% with RMSE of 0.092.