To handle the problems of complicated camera installation, low measurement accuracy and high measurement cost in conventional computer vision measurement methods of fluid surface velocity (FSV) of river models, a new method using local features was proposed. By this method cameras could be installed from a free angle of view without vertical downturn. Firstly throw scraps of paper as tracer particles on the surface. Then rectify the acquired video and reconstruct the aerial view to get the top view of FSV. Next use image matching method of local features to gain the dense velocity vector. Finally optimize FSV and output it. Experimental results show the velocity error of this method is less than 10% when the model velocity is between 0.01 m/s and 0.5 m/s and less than 5% when the model velocity is between 0.5 m/s and 1.5 m/s. In conclusion, this method outperforms the existing ones in terms of measurement accuracy, measurement cost and ease of use, and it has been applied in the analysis and measurement of large river models such as the Changhe section of Xiangjia Dam.