Design and Implementation of Tiny Deep Neural Networks for Landing Pad Detection on UAVs

This paper presents a design paradigm to implement convolutional neural networks (CNNs) on low-power commercial microcontrollers for the detection of landing pads in small-size drone applications. A neural architecture search (NAS) strategy generates and selects CNN architectures automatically; candidate networks are compared in terms of their computing costs and representation capabilities. The proposed NAS procedure adopts a teacher-student learning paradigm, in which the 'student' network should mimic the 'teacher's' intermediate representation. The associate selection strategy aims to attain an efficient feature representation that can take into account the peculiarities of the problem at hand. This approach allowed the generation of tiny networks capable of real-time execution on commercial micro-controllers (STM32 family). Experimental results confirmed that the resulting architectures could trade off generalization capabilities and computing costs effectively, and outperformed state-of-the-art solutions for landing-pad detection in small-size drones.