【摘要】 机器视觉液位测量因其非接触、高精度、体积小、适应多种液体等优势，正成为未来液位检测技术的重要发展方向。然而，其高成本与复杂的部署条件仍限制了实际应用推广。为此，提出一种基于双目立体视觉的透明液体液位测量方法。该方法采用改进的循环全对场变换立体匹配方法(RAFT-Stereo)网络结构，在特征提取阶段引入通道注意力机制，并结合区域加权损失函数以提升透明和弱纹理区域的匹配精度。实验结果表明，改进网络在SynClearDepth数据集上取得了0.53的平均端点误差(EPE)和9.61%的Bad—3.0指标。在透明液体液位高度测量实验中，测量平均误差为1.68 cm,较原始模型降低约78.3%,能够满足实际液位测量的精度需求。更多还原

【Abstract】 Vision-based liquid level measurement offers advantages such as non-contact sensing, high accuracy, compact size, and general applicability across liquid types.However, its high cost and complex deployment requirements limit its widespread practical application.To address this problem, a transparent liquid level measurement method based on binocular stereo vision is proposed.An improved recurrent all-pairs field transforms-Stereo(RAFT-Stereo)network architecture is employed, which integrates a channel attention mechanism during feature extraction and introduces a region weighted loss function to enhance disparity accuracy in transparent and weakly textured regions.Experimental results show that the proposed method achieves an end-point error(EPE)of 0.53 and a Bad—3.0 of 9.61 % on the SynClearDepth dataset.In the transparent liquid level height measurement experiment, the average measurement error is 1.68 cm, which is about 78.3 % lower than that of the original model, and can meet the accuracy requirements of practical liquid level measurement.更多还原