SafeRoute Analytics: A Spatio-Temporal Accident Mitigation Suite

Data Collection & Preprocessing Lead | UNT INFO 5810

Summer 2025 · 5-person team project

Project Summary

SafeRoute Analytics is a decision-ready geospatial intelligence system built by a 5-person team to convert 7.7 million U.S. traffic accident records (Feb 2016 – Mar 2023) into actionable insight for insurance providers, logistics firms, and urban planners. Instead of optimizing only for speed, SafeRoute focuses on safety — using clustering and severity modeling to identify high-liability zones and support lower-risk routing decisions.

Team: Mary Dallas (Introduction & Background), Juayl Bukhari (Related Work Review), Syam Sai Konakalla (Methodology & Analytical Framework), Breanna Kotary (Experiment, Data Analysis & Modeling), and me leading Data Collection & Preprocessing.

My role was owning the data engineering that made the rest of the analysis possible: cleaning and structuring 7.7M raw records down to a modeling-ready dataset, in close collaboration with the team on feature design and clustering.

💼 The Business Problem & Opportunity

• The Problem: Traffic accidents cause unpredictable liability for insurance companies and costly delivery delays for logistics firms.
• The Opportunity: A business-driven model can quantify risk based on real-time variables (weather, infrastructure density, time of day) to inform premium pricing and logistical rerouting.

⚙️ My Contribution: Data Engineering at Scale

I owned the pipeline that turned raw, noisy public data into a modeling-ready dataset:

• High-Scale Engineering: Ingested and cleaned 7.7M records across 49 states, cutting the dataset down to 4.6M rows of reliable signal after quality control.
• Quality Control: Ran missing-value audits per column, dropped features with extremely high (44%+) missing rates, imputed the rest (mean for numeric, mode for categorical), and applied IQR-based outlier removal on environmental variables (temperature, wind chill, pressure, visibility, wind speed, precipitation).
• Feature Engineering: Converted raw timestamps into commute-relevant features (Start_Hour, Start_DayOfWeek, Start_Month, Is_Weekend, accident_duration) that the team’s models used downstream.

Severity Modeling (Team Effort)

The team evaluated Random Forest and XGBoost classifiers on the cleaned dataset (70/15/15 train/val/test split). Random Forest was the more reliable performer, reaching 66.3% accuracy and 65.7% F1, with both models hitting 90% recall on low-severity incidents (Class 0) — strong enough for high-volume logistics filtering, though mid-severity classes remained harder to separate.

🔍 Key Business Insights & ROI

I. Risk-Prone Temporal Windows

Analysis revealed that 8 AM and 5 PM on weekdays are the highest-risk periods.

• Action: Logistics companies can use this to adjust driver shifts or implement “safety surcharges” during peak-risk windows.

II. Infrastructure & Liability (The “Signal Ratio”)

Using DBSCAN and K-Means clustering, the team identified that areas with lower traffic signalization ratios often see higher severity (Avg. Severity 2.64 in the highest-risk cluster).

• Action: City planners can prioritize infrastructure spend (signals/signage) in these specific low-signal/high-severity zones to reduce public liability.

🎨 The Product: “Accident Precaution Assistant”

The final output is an interactive Streamlit/Tableau Dashboard designed for real-time operational use:

• Risk Filtering: Filter hazards by City, Day, and Hour to identify historical “hazardous corridors.”
• AI Safety Advisor: An integrated LLM-powered assistant that converts model outputs into natural language driving tips for fleet drivers.

🚀 Strategic Recommendations

1. For Insurance: Implement dynamic pricing in High-Severity Clusters (36 and 14).
2. For Logistics: Embed the Accident Precaution Assistant API into GPS systems to provide “Hazardous Corridor” alerts during morning rush hours.

View the Full Capstone Paper