Automatic recognition of spurious surface in building exterior survey

Abstract

Buildings consume around 40% of overall energy in the world. Planar mirror detection problem (PMDP) arises when surveying reflective building surface for building energy retrofit. PMDP is also important for collision avoidance when robots navigate close to highly reflective glassy walls. Our approach uses two views from an on-board camera. First, we derive geometric constraints for corresponding real-virtual features across two views. The constraints include 1) the mirror normal as a function of vanishing points of lines connecting the real-virtual feature point pairs and 2) the mirror depth in a closed form format derived from a mirror plane induced homography. Based on the geometric constraints, we employ a random sample consensus framework and an affine scale-invariant feature transform to develop a robust mirror detection algorithm. We have implemented the algorithm and tested it under both in-lab and field settings. The algorithm has achieved an overall detection accuracy rate of 91.0%.