Criação de um dataset para a detecção aérea de pessoas em ambientes aquáticos

Abstract

The increase in drowning incidents in coastal regions highlights the need for technological solutions capable of improving the efficiency and response time of search and rescue operations. In this context, this work presents the development and validation of an automatic system for detecting people in aquatic environments, based on computer vision techniques applied to aerial images captured by unmanned aerial vehicles equipped with optical (RGB) and thermal sensors. A multispectral dataset was constructed, composed of synchronized images acquired at different flight altitudes and lighting conditions, covering real scenarios from the coast of the state of Paraná, Brazil. The images were manually annotated using a single class, person, in order to represent typical risk situations in aquatic rescue operations. To increase data variability and reduce the risk of overfitting, data augmentation techniques were applied, including rotations, horizontal and vertical flipping, scale variations, brightness and contrast adjustments, noise addition, and small translations, in order to simulate different capture conditions, illumination, and target positioning. The experiments were conducted using models from the YOLO (You Only Look Once) family, in different configurations of depth and complexity. A preliminary evaluation of the dataset was carried out, and the models were selected due to their high computational efficiency and suitability for real-time and embedded applications. Different model configurations were evaluated in terms of detection accuracy, robustness to scale variations and spectral modality, and computational performance. The results demonstrate that the combined use of RGB and thermal images significantly improves the reliable detection of people, especially under unfavorable lighting conditions and low visual contrast, which are common in aquatic environments. It is concluded that the proposed approach is technically feasible and presents strong potential for integration into real-world support systems for search and rescue teams, contributing to reduced response time and increased chances of successful rescue.