Object Perception in the Blind Spots with Mirror Based on Depth Prediction Using CNN

Abstract

Object perception in the occuluded region is required for mobile robots, and it may be effective to use mirrors placed at the intersection, in the elevator and so on in order to see something in the blind spots. However, mirrors are difficult to recognize their 3D position by sensing due to high reflectivity. This paper proposes a method of predicting pixel-wise depth of mirror by using CNN. It also proposes a method of flipping object point cloud obtained from sensors along detected mirror plane, which renders virtual objects inside the mirror in real space. By using our point cloud flipping method, robots can percept a person standing at the blind corner.