RoadAI

The page
Information about every road anomaly
Location of where anomalies were reported
Key for the map and filter to see what you want to see
UI
Cool Down to ensure that data doesn't get collected twice
AI classifying the live feed shown

🛣️ About the RoadAI Project

Inspiration

Our inspiration sprang from the chronic and costly problem of deteriorating urban infrastructure. We observed that traditional road maintenance processes—reliant on sporadic human patrols or subjective citizen reports—were inherently slow, inefficient, and reactive. This delay allows small defects, like an initial pothole, to rapidly escalate into major, expensive damage, such as widespread "Alligator Cracking." We recognized the critical need for an objective, automated, and real-time system to capture pavement data and transform it into an immediate, prioritized maintenance plan, focusing specifically on the needs of communities like College Station.

What it Does

RoadAI provides an end-to-end, data-driven solution for municipal road management:

Real-Time Detection: It continuously monitors road surfaces using images captured from webcams or dedicated vehicle-mounted cameras.
Intelligent Classification: It automatically takes a photo of road damage and classifies it based on three criteria:
1. Damage Type (e.g., Pothole, Alligator Cracking).
2. Severity (e.g., Severe, Moderate).
3. Location (Geospatial coordinates and associated road name).
Maintenance Prioritization: It calculates a Priority Score (as seen in the log) by weighting the severity, estimated repair cost, and potentially traffic volume.
Actionable Data: It puts classified anomalies into a comprehensive Anomaly Log and plots them on an interactive map, providing a quick, visual overview of damage hotspots and the Est. Cost for immediate budgeting and resource allocation.

How We Built It

Data Ingestion Pipeline: We established a pipeline to continuously feed images from the simulated Webcam (TM) source into our processing system, tagging each image with accurate GPS data.
Machine Learning Model: We developed and trained a deep Convolutional Neural Network (CNN) for image segmentation and classification. This model was trained on thousands of labeled road images to accurately identify and classify multiple types of damage and assign a severity level.
Cost and Prioritization Engine: We integrated a logistics engine that leverages the model's output (Type and Severity) to look up or calculate two key metrics:
- Estimated Cost: Based on standardized municipal repair rates (e.g., \$1,200 for a Severe Pothole).
- Priority Score: Calculated using a weighted formula.
Dashboard Visualization: The final output is routed to a user-friendly, responsive web dashboard built with a mapping library (like Leaflet or Mapbox), allowing users to view the Prioritized Damage Locations visually and manage the Anomaly Log.

Challenges We Ran Into

Environmental Variability: The model struggled initially with classifying damage under diverse real-world conditions, particularly low light, wet surfaces, and strong shadows, which often obscured fine cracking details. We mitigated this by rigorously augmenting our training data.
Geospatial Accuracy (The "Where"): Ensuring the real-time location tagging was precise enough to map the anomaly to the correct lane or segment of a road, especially when processing images from moving vehicles, proved challenging and required robust error correction in the GPS data.
Computational Load for Real-Time: Running a high-accuracy CNN inference every few seconds demanded significant computational resources. We had to optimize the model architecture for inference speed to maintain low latency, a crucial factor for a "real-time" system.

Accomplishments That We're Proud Of

Integrated Prioritization: Successfully creating a single, quantifiable Priority Score that combines both machine learning output (Severity) and financial data (Est. Cost). This score instantly tells maintenance crews where to go and what to fix first, which is the ultimate goal of the project.
High Accuracy Classification: Achieving a high degree of confidence in distinguishing subtle yet critical damage types, such as separating "Pothole (Initial)" from a simple surface scrape, which directly impacts the severity and cost estimations.
Functional Dashboard: Delivering a clean, interactive dashboard that allows filtering anomalies by Severity Level and Min. Estimated Cost, instantly transforming raw data into practical budgeting and deployment tools.

What We Learned

We learned that developing a civic infrastructure solution requires translating raw technical predictions into business intelligence. The raw "Severe" classification is less useful than the \$1,200 Est. Cost and the Priority 95 score. We gained a deeper appreciation for the need to integrate financial modeling and logistical parameters into core AI output to achieve genuine utility for end-users in municipal government.

What's Next for RoadAI

Predictive Maintenance: Integrate historical data to predict which roads or segments are most likely to develop severe damage within the next 3-6 months.
Integration with Work Order Systems: Develop an API to automatically generate a maintenance Work Order in city systems (e.g., Maximo, Hansen) as soon as an anomaly hits a threshold Priority Score.
Mobile Reporting: Create a simple mobile app for field inspectors to quickly validate AI-detected anomalies and provide feedback to further refine the model.

Would you like me to focus on a specific technical aspect, such as the details of the CNN model architecture or the exact components of the cost estimation model?