PetWatch

Ever stepped away from your desk for two minutes, only to return to find your cat has typed random characters in your workspace? Yeah, me too. My cat thinks my table is the perfect place to nap, my mouse is a toy, and other objects are obstacle to be eliminated.

Motion sensors? They trigger on everything. Manual supervision? Not exactly practical when you're not around. What I needed was something smarter-a system that could tell the difference between me grabbing a pen and a cat jumping onto my desk.

So I built PetWatch: an embedded system that uses computer vision to detect motion, AI to identify if it's actually a pet, and automated deterrents (water spray + alarm) to teach them to stay away. The system runs on three devices working together: BeagleBone Black for vision, Raspberry Pi Pico 2W for networking and control, and a detection server running YOLOv5 for accurate AI identification.

Project Goal: Build a system reliable enough to run constantly, smart enough to distinguish pets from humans and capable of threatening the cats.

The system uses a distributed architecture where each device has a specific job. This separation means if WiFi dies, the camera keeps detecting. If the server goes down, the Pico can still receive frames.

Custom Yocto Linux Distribution

Instead of using a bloated off-the-shelf Linux distro, I built a minimal custom distribution using Yocto Project. This allows to simply write image to an SD card, insert it into BBB's SD slot and boot it. After booting up service needed for the project will start automatically.

Bulletproof UART Protocol

Transferring JPEG images over UART isn't trivial-one corrupted byte and the entire image is garbage. I used a frame-based protocol with CRC-16 validation for both headers and data. The Pico uses an interrupt-driven ring buffer to handle incoming data, then a state machine parser validates everything before acceptance. This allows me using UART at high baudrates.

Real-Time AI That Actually Works

Running YOLOv5 on a server instead of edge device means I can use the large model variant for better accuracy. I tried all of the YOLOv5 models but only large was giving acceptable result. The system responds fast enough for real-time deterrence while maintaining high detection accuracy. When a pet is detected, the server annotates the image with bounding boxes and confidence scores, then sends both the alarm signal and email notification.

Smart Water Pump Control

The latching relay module toggles state with brief pulses, like pressing a button. I implemented pulse generation logic that keeps the GPIO pin in Hi-Z mode normally to avoid interference, then sends precisely-timed LOW pulses to toggle the pump on and off. Multiple activation cycles during the alarm sequence ensure maximum deterrent effectiveness and the sound of pump also works well.