Real-time vehicle tracking and traffic metric estimation using YOLOv11 with CUDA optical flow

We present a real-time traffic analysis pipeline that fuses a one-stage detector with CUDA-accelerated pyramidal Lucas–Kanade optical flow, homography-based ground-plane metrology, and light Kalman smoothing to produce lane-level counts, speeds, occupancy, queues, and rule violations from multi-camera video. On top, an Adaptive Flow-Calibrated Traffic Management (AFTM) layer converts these signals into actionable suggestions (green-split reallocation, reversible/priority lanes, incident response) with hysteresis for stability. On commodity GPUs the system processes multiple 1080p streams in real time, achieving 96–98% counting accuracy and 1.9–3.1 km/h speed MAE on calibrated scenes.