This project provides the first systematic, multi-site validation of low-cost community-deployable edge-AI camera sensors integrated with YOLO-based computer vision for pedestrian and vehicle counting in real-world transportation settings. Building on the team's ongoing PacTrans-funded work on automated detection, tracking, counting, and  safety analysis of pedestrians and cyclists, this project extends that  foundation in two important directions: expanding the YOLO-based  pipeline to include vehicle counting, and evaluating model performance  on data from community-deployable sensors currently being tested in  Portland rather than only conventional research video. 

The project will compare three data processing approaches --- the sensor’s native algorithm, off-the-shelf YOLO models, and PSU’s fine-tuned YOLO models using transfer learning from large-scale datasets such as COCO --- against manually validated ground truth data across 3--5 Portland-area sites. A key focus is understanding how camera angle, height, field of view, site complexity, and model version affect missed objects, double counting, and count error, and how much site-specific calibration is needed for reliable deployment. Rather than requiring extensive manual validation at every site, the project will identify practical validation strategies that agencies can apply efficiently.  

The expected outcome is a set of validated methods and evidence-based guidance for using low-cost camera sensors and YOLO-based models for scalable multimodal traffic counting. By partnering with ODOT (see support letter) and  Green Dormer Solutions, LLC --- developer of the TrafficMonitor.ai/Roadway Biome platform --- the project will also demonstrate how university-developed computer vision methods can transfer into deployable tools for real-world transportation practice,  supporting agencies and communities that need affordable, transparent, and repeatable count data for safety analysis, planning, and infrastructure evaluation.