Helping drones find a safe place to land

My MSc dissertation: combining depth cues with ordinary camera images so a drone can spot safe landing zones in cluttered places, in real time on embedded hardware. Scored 85/100.

The question

A drone’s camera sees colour, not distance. In a cluttered scene, a flat-looking patch of ground can hide a slope, a step or an obstacle. Could depth estimated from that same single camera make landing-zone detection safer, while still running fast enough on the drone’s own hardware?

What I did

My dissertation, Improving Drone Landing Safety with Monocular Depth-Enhanced RGB Segmentation, combined monocular depth estimation with RGB image segmentation to classify safe landing zones. I designed it to reach real-time performance on embedded hardware, not just on a workstation.

Outcome

Scored 85/100, with strong feedback from the external examiner at the defence. It was part of an MSc in Business Analytics that I finished with First Class Honours (Highest Distinction) in October 2025.

What I learnt

Most of the hard work was in the data, not the model. Earlier, a side project classifying cats and dogs had stalled at around 70% accuracy until I fixed mislabelled and imbalanced images. That lesson came back here at a much larger scale.