Exploring higher quality point clouds using stereo vision

Abstract

This research investigates the feasibility of creating a useful high-quality point cloud through stereo vision. We will create a homemade stereo camera setup to implement the stereo vision techniques. To generate the point clouds, we will implement traditional stereo vision using our stereo camera setup, PSMNet, and MiDaS, a monocular depth estimate using neural network. We will compare and discuss if stereo vision may provide better depth estimate, and consequently a more accurate representation of the scene in the point cloud, than single camera approaches. Stereo vision with deep learning enhancement will also be explored, such as the state-of-the-art method, PSMNet. A final evaluation will be done to summarise our findings on which method produces the highest quality point cloud that has the least noise as well as the best depth estimate regarding the subject in focus.