基于轻量化稠密光流的明渠视觉测流方法

Open Channel Visual Flow Measurement Method Based on Lightweight Dense Optical Flow

摘要:

针对机器视觉明渠测流方法中的计算复杂度高、硬件算力需求大及环境适应性不足等问题,提出了一种基于轻量化改进稠密光流的明渠流量视觉监测方法。 该方法通过引入稀疏网格采样策略,在保留关键纹理特征的基础上降低数据维度,从而减少光流计算量。 同时,构建了一种基于垂直积分投影与拓扑约束的智能感兴趣区域提取算法,结合内侧优先策略实现真实水面边界的自动识别与亚像素拟合,提高复杂环境下水面区域定位的准确性与算法鲁棒性。 为验证所述方法的有效性,搭建了明渠流量视觉监测试验系统,构建了 0. 18 ~ 0. 60 m / s 流速区间及 0. 013 9 ~ 0. 044 4 m3 / s 流量区间的数据集。 试验结果表明,在流速测量方面,改进算法的平均相对误差为 3. 44% , 决定系数 R2 达到 0. 935 9,均方根误差为 0. 016 5 m / s;在流量测量方面,系统综合测量误差控制在 3. 42% ,决定系数达到 0. 987 4,均方根误差为 0. 001 1 m3 / s。 结果表明,该方法在保证测量精度的同时,有效降低算法复杂度,具备较好的实时性与工程应用潜力。 本文研究为明渠流量的非接触式、低成本、全天候监测提供了一种可行方案,也为经典稠密光流算法在边缘端设备上的轻量化部署提供了参考。

Abstract:

A lightweight visual monitoring method based on an improved dense optical flow algorithm was proposed to address the limitations of existing machine vision methods for open-channel flow measurement, including high computational complexity, demanding hardware requirements, and limited robustness to environmental interference. A sparse grid sampling strategy was adopted to reduce data dimensionality while preserving essential texture information, thereby lowering the computational cost of optical flow estimation. Meanwhile, an intelligent region-of-interest extraction method based on vertical integral projection and topological constraints was developed. By incorporating an inner-side priority strategy, automatic detection and subpixel fitting of the true water surface boundary were achieved, which improved the accuracy of water surface localization and the robustness of the method under complex environmental conditions. To validate the proposed method, an open-channel visual flow monitoring system was established, and a dataset with flow velocities ranging from 0. 18 m / s to 0. 60 m / s and discharge rates ranging from 0. 013 9 m3 / s to 0. 044 4 m3 / s was constructed. Experimental results showed that the proposed method achieved a mean relative error of 3. 44% , an R2 of 0. 935 9, and an RMSE of 0. 016 5 m / s in flow velocity measurement. In discharge measurement, the overall measurement error was 3. 42% , the R2 reached 0. 987 4, and the RMSE was 0. 001 1 m3 / s. The results demonstrated that the proposed method can effectively reduce algorithmic complexity while maintaining high measurement accuracy, indicating good real-time performance and strong potential for engineering applications. The proposed method can provide a feasible solution for non-contact, low-cost, and all-weather monitoring of open-channel flow and offer a practical reference for the lightweight deployment of classical dense optical flow algorithms on edge devices.