[Paper Review] Intelligent Pothole Detection and Road Condition Assessment

Umang Bhatt, Shouvik Mani+2 more|arXiv (Cornell University)|Oct 6, 2017

Infrastructure Maintenance and MonitoringEngineering3 references21 citations

TL;DR

This paper presents a real-time mobile system that uses smartphone accelerometer and gyroscope data to detect potholes and classify road conditions with 92% accuracy for potholes and 93% for road quality. By leveraging crowdsourced sensor data, the system generates detailed city-wide road condition maps to guide infrastructure maintenance.

ABSTRACT

Poor road conditions are a public nuisance, causing passenger discomfort, damage to vehicles, and accidents. In the U.S., road-related conditions are a factor in 22,000 of the 42,000 traffic fatalities each year. Although we often complain about bad roads, we have no way to detect or report them at scale. To address this issue, we developed a system to detect potholes and assess road conditions in real-time. Our solution is a mobile application that captures data on a car's movement from gyroscope and accelerometer sensors in the phone. To assess roads using this sensor data, we trained SVM models to classify road conditions with 93% accuracy and potholes with 92% accuracy, beating the base rate for both problems. As the user drives, the models use the sensor data to classify whether the road is good or bad, and whether it contains potholes. Then, the classification results are used to create data-rich maps that illustrate road conditions across the city. Our system will empower civic officials to identify and repair damaged roads which inconvenience passengers and cause accidents. This paper details our data science process for collecting training data on real roads, transforming noisy sensor data into useful signals, training and evaluating machine learning models, and deploying those models to production through a real-time classification app. It also highlights how cities can use our system to crowdsource data and deliver road repair resources to areas in need.

Motivation & Objective

• To address the lack of scalable methods for detecting and reporting potholes and poor road conditions.
• To develop a real-time mobile application that uses smartphone sensors to classify road conditions and detect potholes.
• To create data-rich, crowdsource-driven road condition maps for civic authorities to prioritize repairs.
• To improve urban infrastructure maintenance by providing accurate, scalable, and low-cost detection of road degradation.

Proposed method

• Collected training data using two iPhone 6S devices mounted in a 2007 Toyota Prius, one for sensor logging and one for pothole annotation.
• Captured five sensor readings per second, including accelerometer (x, y, z), gyroscope (x, y, z), GPS, and speed data.
• Transformed raw sensor data into meaningful features using time-interval aggregation (10-second intervals for road condition, 2-second for potholes).
• Trained Support Vector Machine (SVM) models on engineered features to classify road conditions (good/bad) and potholes (yes/no).
• Used a dual-app architecture to separate data collection from annotation, reducing data loss and improving reliability.
• Generated city-scale road condition maps by aggregating real-time classification results from multiple users.

Experimental results

Research questions

• RQ1 Can smartphone accelerometer and gyroscope data reliably distinguish between good and bad road conditions in real time?
• RQ2 Can a machine learning model trained on sensor data detect individual potholes with high accuracy?
• RQ3 How can crowdsourced sensor data from mobile devices be transformed into actionable, spatially accurate road condition maps?
• RQ4 What preprocessing and feature engineering techniques improve classification performance on noisy, real-world sensor data?

Key findings

• The SVM model achieved 93% accuracy in classifying road conditions as good or bad, significantly outperforming the base rate.
• The pothole detection model reached 92% accuracy, demonstrating strong performance on real-world driving data.
• Aggregating sensor data over 10-second intervals for road condition and 2-second intervals for potholes improved model interpretability and performance.
• The system successfully generated detailed, real-time road condition maps, such as the Pittsburgh map shown in Figure 11, illustrating spatial variation in road quality.
• The dual-app data collection approach reduced data loss and improved annotation reliability compared to single-device solutions.
• The project demonstrated that smartphone sensors can serve as a scalable, low-cost alternative to traditional road inspection methods.