Low-Power Deep Learning Edge Computing Platform for Resource Constrained Lightweight Compact UAVs

Unmanned Aerial Vehicles (UAVs), which can operate autonomously in dynamic and complex environments, are becoming increasingly common. Deep learning techniques for motion control have recently taken a major qualitative step since vision-based inference tasks can be executed directly on edge. The goal is to fully integrate the machine learning (ML) element into small UAVs. However, given the limited payload capacity and energy available on small UAVs, integrating computing resources sufficient to host ML and vehicle control functions is still challenging. This paper presents a modular and generic system that can control the UAV by evaluating vision-based ML tasks directly inside the resource-constrained UAV. Two different vision-based navigation configurations were tested and demonstrated. The first configuration implements an autonomous landing site detection system, tested with two models based on LeNet-5 and MobileNetV2, respectively. This allows the UAV to change its planned path accordingly and approach the target to land. Moreover, a model for people detection based on a custom MobileNetV2 network was evaluated in the second configuration. Finally, the execution time and power consumption were measured and compared with a cloud computing approach. The results show the ability of the developed system to dynamically react to the environment to provide the necessary maneuver after detecting the target exploiting only the constrained computational resources of the UAV controller. Furthermore, we demonstrated that moving to the edge, instead of using cloud computing inference, decreases the energy requirement of the system without reducing the quality of service.