What is it about?

This research looks at using small, low-power computers like Raspberry Pi and Orange Pi to run artificial intelligence algorithms that can detect objects in infrared images taken by drones. The study tests different sizes of a popular detection algorithm called YOLO to see how fast it can process images and how much computing resources like memory, storage, and power it needs on these small devices. The goal is to find ways to do object detection on drones using small onboard computers that have limited power and resources.

Featured Image

Why is it important?

Doing AI tasks like object detection on small devices is important for drones and other technologies where you want intelligent capabilities but have tight limits on size, weight, and power. This research explores the practical realities of running advanced AI on low-power devices for applications like surveillance, search/rescue, infrastructure monitoring, etc. The results provide guidance on how to optimize different hardware, software, and algorithm choices to best fit the capabilities of small onboard computers for drone-based object detection.

Perspectives

Testing YOLO on small single-board computers was an exciting challenge. Configuring the software and measuring the power and performance precisely took time but provided valuable insights. I'm proud we could rigorously explore AI deployment on low-power devices for the growing drone industry. I hope these optimization tips help fellow researchers and developers build intelligent UAV systems within tight resource constraints. There is more work to be done adapting AI for embedded use, and I look forward to driving further progress.

Andrii Polukhin
National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"

Read the Original

This page is a summary of: Edge Intelligence Resource Consumption by UAV-based IR Object Detection, October 2023, ACM (Association for Computing Machinery),
DOI: 10.1145/3607834.3616566.
You can read the full text:

Read

Resources

Contributors

The following have contributed to this page