Heavy snow substantially degrades the quality and reliability of vision-based traffic monitoring, yet most existing perception benchmarks focus on onboard driving datasets, mixed adverse-weather settings, or broad surveillance scenarios rather than fixed roadside CCTV imagery collected before, during, and after a snow event.
This project develops a public benchmark dataset and evaluation framework for vehicle detection using roadside CCTV camera imagery collected before, during, and after heavy snow events, along with an improved YOLO-based vehicle detection method to handle heavy snow conditions. The study targets a sensing setting that is highly relevant to transportation agencies but underrepresented in existing benchmarks: medium-resolution, street-level roadside cameras deployed for routine mobility, safety surveillance, and operational monitoring. It will synthesize the literature on adverse-weather detection, curate and manually annotate a fixed-camera dataset, develop YOLO-based detection pipelines, and rigorously compare out-of-the-box and retrained models across the three snow phases.
The project is expected to produce four primary outputs. First, a publicly available benchmark dataset of manually labeled roadside CCTV images spanning before-, during-, and after-snow conditions, drawn from open-source Maryland DOT traffic cameras and annotated with bounding boxes, class labels, and metadata such as timestamp, camera ID, and snow phase. Second, a rigorous benchmark evaluation quantifying how heavy snow and residual post-snow scene changes affect detection performance, compared across standard metrics including precision, recall, F1-score, and mAP. Third, a retrained YOLO-based model, incorporating snow-specific augmentation, phase-balanced sampling, and domain-adaptive fine-tuning, that improves robustness relative to the default baseline. Fourth, an open evaluation protocol and trained model weights released to support reproducible research.
Universities Involved
Morgan State University
Principal Investigators
Di Yang
Mansoureh Jeihani
Expected Research Outcomes & Impacts
Beyond these direct products, the project represents a transformative step toward more reliable and intelligent transportation monitoring under adverse winter weather. By focusing on fixed roadside CCTV cameras, a practical sensing platform already widely deployed by transportation agencies yet insufficiently studied for severe weather, the study develops new methods for systematically evaluating how detection performance changes across the temporal progression of heavy snow and how retraining can improve robustness.
The findings will document the extent of performance degradation before, during, and after heavy snow and the degree to which data-driven adaptation can recover lost accuracy, contributing to the broader body of knowledge on weather-robust transportation computer vision. For agencies, the results can inform how existing CCTV infrastructure is used for winter operations, incident monitoring, and emergency response; for industry, the public benchmark supports technology transfer; and for communities, more reliable monitoring supports safer, more resilient winter mobility.
Subject Areas
Computer Vision, Vehicle Detection, Adverse Weather, Traffic Monitoring, Benchmark Datasets
