Low-cost Depth Sensors to Real-time Monocular Depth Prediction for Improved Mapping

Real-time mapping and navigation is a challenging task for robotics, which has been significantly mitigated by the introduction of real-time depth sensors. This thesis presents methods for calibrating sensor data and improving alignment accuracy of overlapping 3D models. Additionally, in an effort to complement the ability of the sensor, machine learning approaches were developed that could estimate depth using only the colour information. Several state-of-the-art approaches to estimate depth from colour were created, and novel training strategies used to improve accuracy and speed. The hope is, one day these systems could remove the need for expensive specialised sensors.