The integration of images and kinect depth maps for better quality of 3D surface reconstruction

Abstract

A novel approach is proposed to improve the quality of depth estimation using Kinect sensor. It is well known that the low resolution of Kinect sensor system makes the depth estimation unreliable at object boundaries due to errors of decoding structured light patterns. In this paper, I propose to integrate visible images and Kinect depth maps to classify depth estimates into two categories: reliable and unreliable depth estimates. The quadratic function is used to model local surfaces and its parameters are estimated from reliable depth estimates. Then the unreliable depths will be re-estimated based on the quadratic function. The experiment shows that the quality of reconstructed plant leaf surfaces can be significantly improved by the proposed approach.