Mobile-Phone-Based Artificial Intelligence Package Development and Validation in Large Scale for Maintenance Asset Management

Kuang, Biao; Chen, Jianli · 2024 · ROSA P / Utah Department of Transportation

archive: archived pipeline: cataloged verified

Get this paper ↗ (full text — opens at the source; we link to it, we don't host it)

Summary

This study addresses the inefficiencies and inaccuracies of manual transportation asset inspections by developing and validating a large-scale, mobile-phone-based artificial intelligence (AI) package. While previous research demonstrated the potential of AI for detecting assets like pavement markings and traffic signs, those efforts were limited by small datasets (~1,000 images) and lacked capabilities for object counting and precise geolocation. This project expands the dataset to approximately 5,000 images per asset type and introduces models to quantify object counts and pinpoint locations, aiming to facilitate more frequent and accurate maintenance planning. The researchers collected video data using a mobile phone mounted on a vehicle’s windshield, capturing footage of pavement markings, litter/trash, traffic signs, and guardrails/barriers. They converted these videos into images and annotated them to train four detection models based on the You Only Look Once (YOLO) deep learning architecture. To address the issue of duplicate detections in video sequences, the team developed a counting model using object tracking algorithms. Additionally, they created a geolocation model that synchronizes video frames with phone-based GPS records using time-based interpolation to estimate the latitude and longitude of each detected object. Finally, an interactive visualization interface was built using the Folium Python package to display detected objects, their classifications, timestamps, and cropped images on a map. The results demonstrate high performance across all developed models. The object detection models achieved F1 scores of 0.88 for pavement marking issues, 0.84 for litter and trash, 0.91 for traffic signs, and 0.96 for guardrails and barriers. The geolocation model exhibited high precision, with an average distance error of only 0.27 meters (approximately 0.9 feet) when validated on a test highway route (I-15) in Utah. The counting and visualization tools successfully quantified assets within specific road sections and provided intuitive spatial representations of asset conditions. The significance of this work lies in its ability to automate and scale transportation asset management. By enhancing detection accuracy and adding counting and geolocation functionalities, the AI package enables transportation authorities to conduct comprehensive, data-driven inspections. This technology supports proactive maintenance planning, reduces reliance on labor-intensive manual assessments, and ultimately contributes to improved road safety and infrastructure reliability. The study confirms that mobile-phone-based AI is a viable, scalable solution for monitoring diverse transportation assets in real-world conditions.

Key finding

The developed mobile-phone-based AI models accurately detected transportation assets with F1 scores between 0.84 and 0.96 and achieved precise geolocation with an average error of 0.27 meters.

Methodology

field_study

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via bulk_ingest_rosap on 2026-05-23 (6 acquisition events logged).

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• signage environment