This study aims to propose a novel autonomous water level monitoring system that integrates edge computing and computer vision to achieve accurate, real-time measurements in resource-constrained field environments, eliminating the dependence on continuous network connectivity and high-performance computing infrastructure.
The proposed system is implemented on a Raspberry Pi platform using a multi-stage algorithmic framework. The process begins with a pre-trained ResNet-50 model that identifies whether a water level gauge is present in the captured images. The image then undergoes a series of morphological processing with precise extraction accomplished through a minimum bounding rectangle technique. Subsequent digit isolation uses Canny edge detection combined with Hough line transformation for geometric correction, followed by a dual-projection method for individual character segmentation. The workflow concludes with a convolutional neural network architecture performing accurate digit recognition to determine water levels.
Experimental results demonstrate the system’s effectiveness and robustness. The integrated approach successfully identified and extracted water level gauge under diverse environmental conditions, showing strong capabilities in tilt correction and digit segmentation. The system achieved low measurement error while maintaining stable real-time performance on the embedded platform, validating its practical applicability for automated field monitoring.
This work provides a novel edge-computing solution through hardware-algorithm co-design, enabling flexible and efficient water level monitoring. The system offers significant practical value for environmental data collection by combining cost-effectiveness with reliable performance in challenging field conditions.
