Mohammed Oussama Chakeur
AI & ML Engineer
A healthy chicken moves constantly — foraging, preening, jostling for space. A sick one doesn't. That behavioral change is the earliest signal of disease, and it happens long before any visible symptom appears. Computer vision is how we read it.
What the camera is actually looking for
The Poultrix camera system, powered by a Raspberry Pi Camera V2 module, captures continuous video from inside the coop. The AI model — built on TensorFlow Lite with a MobileNet architecture — processes each frame to track individual birds and score their movement patterns. The key signals are: movement frequency (how often each bird moves), clustering density (whether birds are grouping abnormally in one area), posture detection (drooping wings, hunched stance), and response latency (how quickly a bird reacts to the flock around it).
Why edge processing matters
Running AI inference on the device — rather than sending video to the cloud — is not just a cost decision. It's a reliability decision. In rural Malaysia, internet connectivity is intermittent. An AI model that requires a cloud round-trip to function is an AI model that doesn't work when you need it most. TensorFlow Lite is specifically designed for edge devices: it runs efficiently on the Raspberry Pi 4's CPU, with inference latency under two seconds per analysis cycle.
The confidence score system
Rather than a binary healthy/sick output, the model produces a confidence score between 0 and 1 for each monitored area of the coop. Scores above 0.7 are logged as normal. Scores between 0.4 and 0.7 trigger a warning-level alert. Scores below 0.4 trigger a critical alert that goes out via SMS and WhatsApp immediately. This tiered system means farmers receive fewer false positives — the single biggest reason farmers ignore automated alert systems.
What it can and cannot detect
The current model reliably detects behavioral anomalies associated with Newcastle disease, infectious bronchitis, and heat stress. It does not replace a veterinarian — it tells the farmer something is wrong and prompts them to look closer. That distinction matters. The goal is early escalation, not automated diagnosis.
The combination of behavioral AI and environmental sensors gives Poultrix a detection window that manual monitoring simply cannot match. Most disease outbreaks show behavioral signals 6–12 hours before any physical symptom is visible. That is the window that saves the flock.
Want to see Poultrix in action?
Get in touch with the team for a demo or technical walkthrough.
