Método baseado em imagens para leitura automática de réguas de vertedouro

Abstract

Spillway gates are critical structures in hydroelectric power plants and require careful monitoring whenever they are operated, since the environmental impact of an incorrect maneuver can affect not only aquatic life in rivers and possible upstream and downstream reservoirs, but also critical structures such as bridges, houses along riverbanks, ferries, and even human lives. In this work, an approach is proposed for the automatic inference of the gate opening value, employing deep learning and classical computer vision techniques on frames captured by cameras commonly used for the remote monitoring of these operations. A YOLO11 model, fine-tuned for the detection of the numeric markers and the pointer present in the scene, is used to obtain the information necessary to estimate the curvature of the measuring scale used to determine the gate opening, rectify the image só that the scale is represented vertically and straight, and refine the position of the pointer tip that indicates the measured value. The proposed approach was evaluated on a dataset containing 27,194 images captured under real operating conditions by 24 cameras, covering three different types of measuring scales. This dataset is made freely available for research purposes. The proposed method achieved a median absolute error of 6.23 cm compared with the reference ground-truth annotations of the gate openings. Considering different maximum opening ranges, the median absolute error of 6.24 cm corresponds to approximately 0.54% of the full scale for openings up to 11.5 m, 0.45% for openings up to 14 m, and 0.39% for openings up to 16 m. These results indicate that, in proportional terms, the measurement deviation decreases as the operating range increases, remaining below 0.6% of full scale in all analyzed scenarios, highlighting the potential of the approach to assist operators in performing remote operations through camera-based monitoring systems.