Joint Human Detection and Scene Classification for Flood Search and Rescue Using Multi-task Learning

Flood is one of the disasters that cost numerous lives in the Philippines and in the world, as well. Search and rescue (SAR) operations are usually focused on general disasters and on the task of human detection alone. In this paper, a multitask learning (MTL) model is created to perform joint human detection and scene classification exclusively for flood occurrences. Various detection and classification models were tested to find YOLOv7 and MobileNetV2 to be the most suited for the tasks. Although the MTL model resulted to be 3.16% smaller than the two models combined, its processing time could still be improved. Upon evaluation, the model achieved an AP of 77.17% and an accuracy of 76.00%, which are not far off from other similar MTL works. Overall, this demonstrated promising capabilities for flood SAR applications.