Water surface velocity measurement plays a pivotal role in watershed flood prevention and water resource allocation, providing essential data for timely flood warnings and scientific water management. Particle Image Velocimetry (PIV) is a conventional approach to estimate flow velocity by tracking small floating particles on water surfaces. However, this method frequently encounters practical limitations due to challenging environmental conditions, including surface glare and sparse particle distribution, which often lead to tracking failures. To overcome these limitations, we develop an advanced particle tracking technique that incorporates deep learning algorithms. The proposed method combines a YOLOv5s-based particle detection system with a DeepSort tracking framework. This integrated approach enables real-time velocity monitoring while demonstrating remarkable robustness against common environmental disturbances such as variable illumination and low particle concentrations. Field validation experiments achieved a mean velocity error (MVE) of 0.015 m/s under realistic operating conditions.