Design and Deployment of an Efficient Landing Pad Detector

Robust landing pad detection plays a major role in Autonomous Unmanned Aerial Vehicles (UAVs). This problem can be approached using deep neural networks for vision-based inference. However, the full integration of deep learning algorithms into the small UAVs is still challenging for their limited resources. This paper presents a landing pad detection pipeline based on a revisited version MobileNetV3-Small. The proposed architecture inherits robustness from the general-purpose version but limits the computational cost significantly thanks to a set of design criteria aimed to limit hardware requirements. Experimental results confirm that the proposed network compares favorably with a lightweight general-purpose object detector in terms of accuracy/computational cost trade-off. The system is also deployed on a commercial general-purpose microcomputer confirming that satisfactory performance can be obtained on general-purpose embedded architectures.