Cascade method for water level measurement based on computer vision

Abstract

Computer vision-based methods of water level measurement that utilize cameras to capture and process images of water bodies and their surroundings are gaining attention due to their advantages over non-visual sensors. This study aims to improve the generalization ability of the water level measurement algorithm based on computer vision to promote the application of the method in a broader range of scenarios. First, we briefly introduce a pipeline consisting of two main steps: calibration and measurement. Second, we propose a novel cascade model that comprises global and local subnetworks to achieve a more precise waterline position coarse-to-fine. In the training phase, apart from basic data augmentation methods, we employ a multiscale training approach to utilize samples more effectively. Finally, compared with other methods, this study increases the accuracy rate and showcases superior accuracy, generalization ability, and application potential.