A Solar Raspberry Pi 5 Counts Traffic on Just 5.6 Watts
A student-built Raspberry Pi 5 runs YOLOv8 Nano on solar power, averaging 5.6W and counting vehicles with 96.6% accuracy. No AI accelerator required.
You do not need an AI accelerator, or even a power outlet, to count cars with computer vision. A Raspberry Pi 5 traffic monitor built by University of Cyprus student Marios Christoforou, featured by Hackaday on September 24, 2026, runs a YOLOv8 object detection model on the Pi's own CPU, draws an average of just 5.6 watts, and lives entirely off a solar panel and battery.
- Hardware: Raspberry Pi 5 with 4GB of RAM, a 720p USB webcam, a 100W solar panel and a 600Wh 12V battery
- Model: YOLOv8 Nano running at about 10 frames per second, with no NPU or AI HAT
- Power: 5.6W average, or roughly 134Wh per day
- Accuracy: 96.6% on pre-recorded test footage, including at night under street lights
How Does a Raspberry Pi 5 Count Traffic?
The Pi runs a headless Raspberry Pi OS install and feeds webcam frames to YOLOv8 Nano, the smallest version of the popular YOLO detection model. It classifies cars, trucks, buses and motorcycles, tracks which direction each one is travelling, and uses logic to avoid counting the same vehicle twice as it crosses the frame. Every count is uploaded to a MySQL database with a UTC timestamp, so the data is ready for charts and analysis.
The 96.6% accuracy figure comes from standard pre-recorded test footage, and the builder reports that the system keeps working after dark under ordinary street lighting. For a single-board computer running a vision model on its CPU, that is an impressive result.
The Low-Power Tricks That Make Solar Work
The clever part of this Raspberry Pi project is the power budget. A Pi 5 running a neural network flat out would drain a battery quickly, so the build trims everything it does not need:
- Radios and ports off: Bluetooth, PCIe, audio and HDMI are disabled
- Underclocking: the CPU and GPU run below stock speed, trading headroom for efficiency
- Smarter frames: unimportant regions of each frame are discarded before detection
The result averages 5.6W. The 600Wh battery was sized to keep the counter running for at least three cloudy winter days, with the 100W panel topping it up whenever the sun is out. That makes it deployable on a roadside pole with no mains power.
No AI Accelerator Needed
It is worth pausing on what this project does not use. Many edge AI builds reach for a Hailo HAT or a USB accelerator, and those are great options, as our AI accelerator stick comparison shows. But a careful choice of model and power settings let the Pi 5's CPU carry the whole job on its own, the same lesson we saw in running local LLMs on a GPU-less server.
Why This Maker Project Matters
Traffic counting is expensive when done with dedicated commercial hardware, and that cost keeps a lot of smaller towns and roads unmeasured. A solar Raspberry Pi 5 with a webcam brings that data within reach of a student budget. For more builds like this, explore our Raspberry Pi and mini computer coverage.
Sources: Hackaday — September 24, 2026; Marios Christoforou project page — project write-up.
More Mini Computers Stories
MaTouch ESP32-S3 E Ink Board: A $60 Voice-Ready Dashboard
Makerfabs' MaTouch ESP32-S3 pairs a 7.5-inch four-color E Ink screen with a four-mic array for $59.80. Here are the specs and what you can build.
ASRock NUC 300 Mini PCs: Wildcat Lake With Dual 2.5GbE
ASRock Industrial's NUC 300 series brings Intel Core 5 320 and Core 3 304 to mini PCs and boards with up to 64GB of DDR5-6400 memory and dual 2.5GbE.
DGX Spark 64GB: Who NVIDIA's $4,999 Local AI Box Is For
NVIDIA's DGX Spark 64GB starts at $4,999 on Oct 23 and runs models up to 100B parameters. Here is who the local AI box suits and how clustering works.