Deep Learning Implementation on Aerial Flood Victim Detection System

Abstract

Hydrometeorological hazard such as floods are considered as a regular natural disaster in Indonesia due to its frequent occurrence. To mitigate the risk, search and rescue operations need to be carried out immediately. The sheer magnitude of floods poses a major challenge for responders, and the emerging drone technology could help to alleviate the problem due to its deployment speed and coverage. Automation in drone technology has potential to improve its effectiveness. This paper explores the idea of human detection during floods using a computer vision approach. This approach utilizes a one stage detector model as detection speed is crucial in disaster management case. The dataset used for training consists of 200 labelled and negative images taken from drone point of view. This paper conducted 3 experiments to find out the difference in performance when the model was trained on flood and non-flood dataset, as well as the effect of image input size to the model’s performance. The first experiment was trained on non-flood dataset. The second experiment was trained on flood dataset, and the third experiment is the modified version of the second model. The results show that the model trained on flood dataset performed worse than non-flood counterparts with the non-flood mAP reached 90.80% while flood mAP reached 39.15%. In addition, the experiments also conclude that increasing the input size of image during training, will increase the detection performance of the model at the cost of FPS