Abstract Manager

Accurate identification of curbside parking restrictions is essential for estimating effective on-street parking availability and for supporting the design of residential parking allocation programs. While recent advances in computer vision (CV) have enabled automated traffic signage detection, most existing models are typically trained on generic or non–North American datasets and are not tailored to parking-relevant signage or the operational needs of municipal curbside management. As a result, signage-related constraints, such as no-parking zones, seasonal parking zones, accessibility parking, fire hydrants, and bus stops, are often excluded from parking occupancy analyses, leading to overestimation of available curbside space.

This study develops and evaluates a mobile vision–GPS framework for detecting, classifying, and spatially tracking traffic and parking-related signage relevant to the City of Edmonton’s Residential Parking Program. High-resolution 4K video data are collected using a vehicle-mounted GoPro camera oriented to a front view at approximately 135 degrees relative to the roadway, enabling continuous capture of roadside assets under real-world operating conditions. Detected signage includes parking and no-parking signs, accessibility parking, fire hydrants, bus stops, crosswalk-related signage, speed bumps, stop signs, and yield signs, each of which directly influences curbside parking eligibility.

A systematic comparative analysis is conducted across multiple deep learning–based object detection models, including YOLOv8 and YOLOv26 variants, to evaluate detection and classification performance. The study examines the influence of model architecture (large vs. extra-large), training strategies, dataset splits, image resolution, batch size, learning rate schedules, confidence thresholds, and mosaic augmentation settings. Model performance is assessed using standard detection metrics, including precision, recall, F1-score, and mean Average Precision, evaluated both overall and per class.

The results will assess performance variations across CV models’ architectures and hyperparameter configurations, with clear implications for detecting smaller, less frequent, or visually complex signage classes common in North American residential contexts. By identifying optimal model configurations and quantifying trade-offs between accuracy and robustness, this work aims to provide a validated detection framework that supports curbside space estimation, asset mapping, and parking policy enforcement. The proposed approach is scalable and transferable, offering municipalities a practical pathway for integrating vision-based signage analytics into data-driven curbside management systems.