GeoAI Behavioural Pattern Recognition for Smart Cities
The challenge
Pet owners demand reliable lost-pet detection and health anomaly alerts without sacrificing battery life or increasing cost. Traditional geofencing produces too many false alerts, as crucial behavioural context is missing. This required a deep investment in product engineering to deliver a high-value, battery-efficient solution for consumers.
What we built
We designed the Felica Pet Tracker as a hybrid BLE + GNSS collar with an IMU. Its core is an AI-powered embedded system that runs a tiny ML inference engine (TensorFlow Lite Micro / Edge Impulse) onboard. This Edge AI embedded system classifies behaviour (resting, roaming, erratic escape behaviour), increasing GNSS duty-cycle only when an enriched "escape" event is detected. This strategy improves detection sensitivity while dramatically reducing useless GNSS wakes.
How it works
The embedded systems development focused on an IMU that sleeps with FIFO gated, waking the MCU only on threshold crossing. ML models, trained on instrumented canine/cat datasets, were compressed via quantisation to fit 64-128 KB RAM devices—a hallmark of expertise. Server-side, our models fuse historical tracks, owner-provided "known-safe areas," and social-collar sightings to produce a ranked recovery route using GeoAI. This intelligence enables Predictive maintenance AI IoT by providing alerts based on behavioural anomalies, not just simple location.
Results
Pilots showed a reduction of false "missing" notifications by ~72% versus geo-fence only, and median time-to-recover improved by ~45% when using the ranked recovery routes.
Scalability & Security: Typical battery life is 2-6 weeks for frequent users, with the potential for 3-6 months for low-activity pets. Provisioning uses NFC to write secure keys to the secure element. This device serves as a micro-asset in a future Digital twin smart city, proving that highly optimized IoT technology can be widely deployed for public safety and premium services.