UAV embedded system - a selection process

Deep Neural Networks are considered a reference point in many applications, but they come with a higher cost related to the system used resources, i.e., memory and computational power. To build such real-time embedded systems is a challenge and a difficult choice, at least from the point of view of the best-fitted system like the one that should be specifically chosen for the computer vision tasks, to give an example. In this paper, such a task represents the central element of several different tested systems used to detect humans from a quadcopter. In this respect, a comparative study between detection systems implemented with one of the most general embedded platforms (i.e., a Raspberry Pi platform sustained or not by a neural engine Neural Compute Stick 2) and, respectively, with the Jetson Nano, as a dedicated embedded platform, is conducted and the results are discussed. Due to the specific characteristics of each embedded system, due to the particular hardware and software optimizations, different performances were obtained, even for similar detection systems. All these illustrate the impossibility of choosing the optimal system for a certain application only based on the catalog data analysis of both processor and system as a whole. The paper also highlights the advantages and limitations associated with each of the analyzed embedded system, specifically from the perspective of the functionality provided to the final detection system.