CAMRA: Context-Aware Road Risk Assessment System

Inspiration

Road anomalies such as potholes and damaged surfaces are common causes of unsafe driving conditions, vehicle damage, and increasing road-maintenance costs. Conventional inspection methods are largely manual, periodic, and difficult to scale. This motivated us to develop a low-cost Edge AI system that can continuously monitor road conditions and provide useful information beyond simply detecting a pothole.

What We Built

We developed CAMRA (Context-Aware Road Risk Assessment System), an Edge AI-based road monitoring framework designed to detect road anomalies in real time and assess their practical risk.

The system uses a camera connected to a Raspberry Pi 5 (8 GB) to capture road video. A lightweight YOLO26 detection model performs real-time anomaly detection, while an optimized ONNX/NCNN inference pipeline enables efficient CPU-based execution on the Raspberry Pi.

Instead of treating object detection as the final output, CAMRA introduces a contextual assessment layer. Confirmed detections are analysed using factors such as detection confidence, apparent damage size, temporal stability, road impact, observation persistence, and spatial consistency. These factors are used to generate a meaningful road-risk assessment and maintenance priority.

The overall workflow is:

$$ \text{Camera Input} \rightarrow \text{YOLO26 Detection} \rightarrow \text{Temporal Verification} \rightarrow \text{Context Analysis} \rightarrow \text{CAMRA Risk Assessment} \rightarrow \text{Maintenance Insight} $$

What We Learned

Through the development of the project, we gained practical experience in computer vision, Edge AI, deep-learning deployment, embedded computing, model optimization, and real-time video processing. We also learned that achieving good detection accuracy is only one part of building a useful real-world system; the output must also be interpretable and relevant to the final application.

How We Built It

The system was developed as a modular pipeline. The Raspberry Pi 5 handles camera acquisition, preprocessing, model inference, contextual processing, and event logging. YOLO26 provides the road-anomaly detection layer, while ONNX and NCNN are used to prepare the model for efficient embedded inference.

The architecture was designed so that the detection model and the CAMRA decision layer remain modular. This allows future improvements to the detection model or additional contextual information to be incorporated without redesigning the complete system.

Challenges

The major challenges included achieving real-time inference on CPU-based edge hardware, balancing detection accuracy with computational efficiency, handling false detections caused by lighting and motion, and designing a meaningful assessment layer beyond conventional object detection.

These challenges led us to focus on a lightweight and explainable approach in which visual detection is combined with temporal and contextual information.

Our Goal

CAMRA aims to demonstrate how a low-cost Edge AI platform can move from "detecting a road defect" toward "understanding its potential road risk." The long-term vision is to support data-driven road maintenance, improve infrastructure monitoring, and provide a foundation for intelligent transportation and smart-city applications.