Abstract: The wave represents the motion of river flow. The surface speed of river can be estimated through motion analysis for the wave. In the paper, a method was proposed based on computer vision to estimate the surface speed of river directly. The method tried to capture the motion of wave caused by flowing river from the video. However, even taken by HD camera, the contrast between moving waves and even surface in an image is still not obvious since they are homogeneous and all moved as a whole. In order to enlarge details of the motion of waves, the map of motion saliency was calculated by the way of frame difference method. In the map, the key points were extracted and characterized by SURF features. These key points represented the most salient positions of waves. Through the point matching algorithm, a key point in one map and its counterparts in next map were searched. The correspondence between the 2 matched points indicated the motion of wave in the video and the distance between them was computed. In principle, with this distance and the parameters of camera, we could estimate the immediate speed of flow. However, the distance was noisy essentially. For robust and accurate estimation, we estimated the average speed instead of immediate speed. So, we calculated the histogram of the distances during the period of time. We found that most of these histograms appeared as uni-modal distribution. However, there existed some histograms which appeared with 2 adjacent peaks, or appeared with a flat peak. This resulted in the difficulty for estimation of distance accurately. To address the problem, we utilized the Gaussian curve to fit the histogram. The peak of the fitted curve could be searched accurately and its corresponding distance was viewed as the optimal estimation of average distance. Finally, with the speed formula derived from pinhole model, the optimal distance and the time between 2 maps, we could estimate the average surface speed of the river flow. To validate the availability of the proposed method, we compared the speeds estimated by our method with the baselines measured by the current meter. In our experimental setting, we selected gently surface for measurement task, without whirlpool and reflection. We conducted 8 measurements, with the speeds being limited between low and middle range. The experimental results showed that maximal relative error of speed between ours and the baseline was 3.12% while the min relative error only 1.39%, indicating good accuracy of our method. The min and max coefficient of variation was 1.04% and 1.63% respectively, showing high reliability. The correlation coefficients of Pearson and Spearman between our estimators and measured values were respectively 0.998 and 0.990. Bland-Altman regression P is 0.16, higher than 0.05 and in Bland-Altman scatter plot, most of points fell into the limits of agreement. These results showed that the flow speed estimated by our method had a good consistency with the baselines. In addition, our method was compared to the image processing method by previous literatures, the results showed that the time consumption was shortened by our study, which was only 4.4% of that of the literature, indicating that our method is faster than the previous method. In sum, this study provides an effective method for the estimation of flow speed of rivers with complex background.