AI birds detection app
Applying a computer vision model to bird recognition at a feeder
Every day brings news about the rapid development of AI - new models, new GPUs, new devices. At the same time, there is a real need for intentionally simple projects, especially when the goal is to show students how neural networks work and how they can be used not only in industry, but also in education.
For this reason, I designed and implemented a small project for one of the schools in Bratislava, Slovakia. The main goal was to show students how easily ML technologies can be used in their own research.
The task was to collect data on the birds living around the school. For this, we used a small Intel NUC computer connected to a USB camera. The hardware was intentionally simple and inexpensive - the kind of equipment any school can realistically have.
While the hardware was straightforward (“use what the school already has”), the software part required more thought. We needed a neural network that could run on weak hardware. I chose the Ultralytics YOLO model, which can recognize many classes of objects, including birds.
A small Python program was written to capture the video stream from the USB camera, detect birds in the image, and save the results into a dedicated folder - storing both the cropped photo of the bird and a CSV log with the timestamp, detection confidence, and other details.
Everything was installed on a windowsill in the school. The camera was pointed outside, at a small improvised feeder with some seeds and peanuts.
The project started working successfully, and today we received the first photo of our visitor - a magpie attracted by the seeds.
From the technical side, we learned a few things:
- YOLOv8n is an excellent choice when you need object detection on low-power hardware.
- Open Source is amazing: you can get a practical working solution for free.
- CPU inference can easily replace GPU for simple tasks - modern CPUs are more capable than many people expect.
The project was well received by the teachers and will likely be developed further. Our plan is to collect enough images of local birds, identify the species, and train this model (or another one) to classify birds by species.
We’re very happy we were able to launch this project - it shows that you don’t need expensive hardware or the latest GPUs to do meaningful work. What matters far more is investing in education. Students need clear, understandable engineering.