Destination-Based People Counting via Full-Cross Line Analysis Using a CPU-Friendly OpenCV–ONNX Pipeline

This paper presents a lightweight, real-time people counting system that determines entry and exit direction using an OpenCV–ONNX pipeline running entirely on CPU hardware. The system combines YOLOv5s for object detection and geometric reasoning for destination-based counting without relying on GPU acceleration. Two test videos were analyzed (42 s and 56 s), achieving 100% counting accuracy and an average processing speed above 30 FPS. The proposed framework demonstrates that efficient, interpretable, and deployable computer vision systems can be achieved using accessible open-source technologies.