Leveraging Image Semantic Segmentation for Road Infrastructure Localization in Traffic Safety Applications

Abstract

Ensuring road safety is a critical concern in urban planning and transportation management. Current approaches to monitoring roadway assets rely on periodic, labor-intensive nighttime field surveys and lack an efficient way to exploit crowd-sourced street view imagery for real-time insight, leaving agencies without timely information on infrastructure such as street luminaires that are vital for nighttime driving safety. This study explores the application of machine learning (ML) to enhance the development and maintenance of robust roadway and roadside inventories. Utilizing image segmentation techniques applied to Google Maps street view images, the research focuses on detecting and localizing street luminaires, which are crucial for nighttime driving safety. The detection module achieved high precision (≥0.98) and recall (≥0.90), demonstrating its effectiveness. However, the localization module faced challenges in complex scenarios, such as varied camera positions and uncommon objects like steel bridges. Factors influencing performance included the camera’s lane position and the limitations of third-party image databases. Recommendations for addressing these challenges include the use of video logs for accurate location recording, fixed survey cameras for precise distance calculations, and consistent survey protocols. The findings underscore the potential of ML to improve roadway inventory management. Future research should focus on expanding the training dataset, optimizing hyperparameters, and conducting comprehensive tests in diverse environments. By implementing these recommendations, transportation agencies can leverage ML to create more accurate and efficient roadway inventories, enhancing road safety and reliability.