ETD

• edge Ai
• embedded system

This thesis describes the design, testing, and implementation of an energy-aware edge AI
orchestration system using the Arduino UNO Q device for real-world object detection in remote
and wildlife monitoring settings. These settings often face challenges like limited power supply,
changing environmental conditions, and unreliable connectivity, making traditional cloud-based
AI solutions impractical. To tackle these issues, we developed and evaluated a lightweight and
efficient edge-based intelligent system. The first part of the research focuses on testing lightweight
object detection models on the Arduino UNO Q. We evaluated two major architectures, Fast Object
More Object (FOMO) and YOLO-PRO, with different configurations for input resolution and
quantization techniques. We analyzed a total of 30 benchmark models based on metrics such as
inference time, frames per second (FPS), RAM usage, model size, and detection accuracy. The
experimental results showed that FOMO models provided significantly higher processing speeds
due to their lightweight design, while YOLO-PRO models offered better detection accuracy and
object localization performance. Increasing the image resolution improved accuracy, but it also
raised latency and memory consumption. INT8 quantization enhanced inference speed and
reduced memory usage with little effect on accuracy. The study found the best configurations for
speed-focused, accuracy-focused, and balanced deployment scenarios. The second part of the
thesis presents an energy-aware framework for coordinating multiple devices, allowing smart task
assignment across several Arduino UNO Q worker nodes. A central controller monitored each
node’s battery level, charging status, CPU load, workload status, and energy use. Using a health score-driven scheduling algorithm, the controller assigned inference tasks to the most suitable
worker nodes. Weak or overloaded nodes were avoided, while healthier nodes handled more
processing tasks to ensure the system operated continuously and reliably. We implemented the
proposed system with Python, Flask APIs, and the Edge Impulse Linux SDK. It was tested through
various scenarios including worker registration, inference execution, scheduler evaluation, weak node management, and high-load testing. The results showed that dynamic orchestration greatly
improves reliability, energy efficiency, and workload distribution in resource-limited edge AI
settings. Overall, this work demonstrates the potential for deploying intelligent, energy-aware edge
AI systems with low-power embedded hardware for real-world applications.