SnowSense IoT Network

Problem Statement

Accessibility has long been an underaddressed challenge in municipal planning. The winter season compounds these barriers significantly: icy sidewalks, accumulating snow, reduced visibility, and cold-sensitive medical conditions all restrict mobility for people with disabilities, seniors, and other vulnerable residents. Traditional patrol-based inspection models are labour-intensive, reactive rather than preventive, and cannot scale to cover the region’s full pedestrian network in real time.

There is a clear need for a low-cost, data-driven solution that can detect hazardous surface conditions as they develop and communicate that intelligence directly to city operations staff.

How we built it

Embedded Hardware (Arduino)

• HC-SR04 ultrasonic sensor measures distance to surface obstructions in real time, outputting readings every 100 ms
• DHT11 / temperature sensor captures ambient temperature for ice-risk correlation
• LED warning system provides on-device visual alert — green (safe) or red (hazard)
• Servo-based barrier protection for preventing entry into hazardous zones

Real-Time Flask Dashboard

• Leaflet.js map & municipality data sidebar with markers for green (safe), red (hazard) or purple (anomaly)
• Node Status panel: live distance, temperature, and GPS coordinates updated every 3 seconds
• AI Risk Score bar: logistic-regression-style sigmoid classifier scoring 0–100% based on distance, temperature, humidity, and snow depth
• Atmospheric Conditions panel: Open-Meteo API with OpenWeatherMap information
• Live connection indicators for sensor network, map layer, weather API, AI risk model, and LOF anomaly model

Anomaly Detection — Local Outlier Factor

The system implements unsupervised machine learning anomaly detection using scikit-learn’s LocalOutlierFactor. This approach requires no labelled training data and adapts automatically to the evolving baseline of a given sensor node’s environment.

Feature Vector (4-Dimensional)

What Makes This Different

• Combines real-time IoT edge sensing with server-side unsupervised ML
• Zero labelled training data required: LOF adapts to each deployment location’s unique baseline automatically
• Scalable: 16-metre ultrasonic detection range means a single mobile unit can survey an entire sidewalk block in one pass
• Accessibility Outcomes: City operations staff receive real-time hazard maps without waiting for resident 311 complaints or scheduled inspections
• Anomaly detection catches sensor malfunctions before they produce false safe readings, improving reliability for vulnerable users
• Prioritizes snow removal: hazard data can be routed directly to dispatch systems to schedule clearing at curb ramps and crossings first

What's next for SnowSense IoT Network

• Supervised ML upgrade: once labelled data accumulates from LOF-flagged events, train a RandomForestClassifier on [distance, temperature, Δdistance, Δtemperature, hour_of_day, humidity, snow_depth] to replace the Arduino threshold with a server-side model
• Predictive pre-alerts: use Open-Meteo’s hourly forecast to push pre-emptive warnings when the temperature + humidity forecast crosses ice-formation thresholds

Built With

• arduino
• chart.js
• dht11
• flask
• hc-sr04
• html/css/js
• leaflet.js
• open-meteo
• polling
• pyserial
• scikit-learn