Sheridan INFO53883 · Capstone Team 2B · Ontario highway safety

Highway risk you can read before you drive.

Smart-Shield fuses Ontario 511-style alerts (NLP), road-surface vision, and environmental collision risk into one Safety Score S, with speed guidance for Ontario highways and 400-series corridors.

See it in action Try the Safety Score View repository

Live map on Render. On a phone, use /mobile. Free hosting may take about a minute to wake on first open.

About this project

Smart-Shield AI is a Sheridan College team project (Team 2B). Author: Afolabi Adesina.

My role: Team Lead.

Proposed the project topic.
Led the exploratory data analysis methods.
Developed the mathematical formula behind the Safety Score and shaped parts of the analysis to match it.
Recommended OpenStreetMap over Google Maps to avoid API fees.
Trained the model on a DGX server.

Teammates sourced datasets and research papers alongside me and contributed design input.

Why this project: I took this project personally. Working in a manufacturing plant at Lafarge Africa taught me that safety means everyone goes home safe at the end of the day.

Back to the portfolio home

Three model inputs

Each pillar outputs a 0–1 risk signal. Fusion weights produce the operational Safety Score.

Pillar T

NLP Brain

TF-IDF over hazard lexicon from unstructured alerts (ice, collision, closure) → text risk T.

Pillar V

Vision Brain

ResNet18 (+ optional autoencoder) on road imagery → surface hazard confidence V.

Pillar E

Environment

Weather / surface / lighting signals calibrated to literature → environmental index E.

See it in action

Open the live map demo on Render. This is the map-in-action view with routes, the 403 exit demo, and settings. The Safety Score sliders below stay on this page.

Open live map demo Open phone layout (/mobile)

Try this first: tap Play 403 exit demo.

Cold start: free hosting may take about a minute to wake on first open.

What you will find

Find safest route: enter start and destination. Safety ratings: 0–30 low, 31–70 medium, 71–100 high. Turn-by-turn directions are included.
Play 403 exit demo: simulated drive with trip and driver safety scores. Export CSV when you are done.
Settings: road conditions (clear, wet, blizzard, ice storm), day or night, camera alerts, voice guidance, and how camera images are scored.
Data: OpenStreetMap and City of Toronto open data. Not official MTO or 511 guidance.

Interactive Safety Score

Drag the sliders to update S, the risk band and speed advice. Formula: S = (w_T·T + w_V·V + w_E·E) × 100
Notebook defaults: w_T=0.25, w_V=0.35, w_E=0.40

Live fused score 43

Text risk T = 0.35 Vision risk V = 0.40 Environment E = 0.50 Weight w_T = 0.25 Weight w_V = 0.35 Weight w_E = 0.40

Decision table

Moderate Risk

Recommended speed: ~80% of posted limit (e.g. 80 km/h).

S	Category	Speed advice
0–30	Low	~100% posted
31–70	Moderate	~80% posted
71–100	High	~60% posted (freeway advisory floor often shown as 80 km/h)

Latest Validation Snapshot

Values below are synced to the latest notebook outputs and D3 evidence pack. The model was tuned to catch rare high-risk events (recall) at the cost of many false alarms (low precision), which is typical for rare-event safety screening.

Check	Current value	Why it matters
Stage A Fatal recall	1.000 @ threshold 0.5 (precision 0.0008)	KPI gate met for missed-Fatal risk, with false-alarm trade-off disclosed
Best baseline MCC	KNN = 0.3819	Best raw ranking score among 3-class baselines
Deploy model	RF Tuned (MCC 0.1641)	Chosen for SHAP explainability and auditability, not MCC leadership
Geo fairness audit	Parity gap = 0.0387 (PASS ≤ 0.05)	Urban vs Suburban/Rural accuracy gap currently within threshold

How to run the project

Prefer the split notebooks under notebooks/parts/. They are easier to run than one giant Run All.

Part	Notebook	Focus
01	01_charter_eda_features.ipynb	Charter → EDA → features
02	02_vision_brain.ipynb	ResNet + AE (GPU recommended)
03	03_tabular_ml.ipynb	SMOTE, GridSearch, ethics
04	04_fusion_deploy.ipynb	Fusion, SHAP, deploy artifacts

Links and local demo

This hosted page is an interactive Safety Score explorer and documentation. The map demo also runs live on Render (phone layout at /mobile). You can still run the Flask demo on your machine if you want the full local stack.

Repository Source, notebooks, docs, and demo code. File guide PROJECT_FILES.md, with every folder explained. Notebook parts Split pipeline for GPU vision + fusion. Author profile afolabi-adesina-m on GitHub. D3 evidence pack Exported figures and metrics tables.

Optional local map demo (live CCTV + models)

cd demo
pip install -r requirements-demo.txt
python api_server.py
# → http://127.0.0.1:5050
# Real Vision mode: nearest Ontario 511 CCTV → ResNet18 → V in Safety Score S

Prefer the live Render demo for a quick look. This GitHub Pages site does not run Flask, GPU inference, or the model weights. Local setup is for development with the trained models folder.