3D image reconstruction based on current stereo vision techniques

Abstract

3D image reconstruction has become progressively popular in recent years with its application ranging from facial recognition in smart phones, measuring the deformation of an object, and even to reconstructing 3D environments in autonomous vehicles. Most of the 3D image reconstruction techniques is derived based on the principle of stereo vision. Stereo vision is based on the term stereopsis which refer to how our eyes perceive depth. There is an underlying debate in the autonomous vehicle industry between the use of active setup like LiDAR and passive setup with just cameras. LiDAR is popular within the autonomous vehicle industry due to its high accuracy and reliability when detecting objects. However, LiDAR is expensive and bulky to apply in autonomous vehicle. Instead, using a passive stereo setup will be much cheaper, smaller and easier to apply. However, passive stereo setup is susceptible to objects that have weak texture. Thus, we would like to find out to what extend does texture affect the accuracy of the 3D image reconstruction in a passive stereo setup. In this research project we will understand the principle of stereo vision and construct a stereo vision system to examine how textures affect the 3D image reconstruction. This project will look at a traditional stereo vision technique that is implemented using OpenCV and also compare the results we get to a deep learning stereo vision technique.