Drowned victims detector via attention-based U-Net and fuzzy similarity analysis

Finding drowned victims at the bottom of a water reservoir often depends on the analysis of data obtained from various underwater sensors. In this paper, we propose a system based on the study of sonar data, where the acquired image is automatically divided into smaller fragments and subjected to rapid segmentation to detect a potential body. For this purpose, the U-Net network architecture was modeled, considering different feature techniques. The model was extended to use a spatial attention mechanism to increase the model’s effectiveness during training. The output is processed by contour normalization, and analyzed by comparison technique using Hu Moments. The proposed system was tested on data obtained in Szczecin, Poland, where the system was looking for a dropped effigy to the bottom.