Joint Disaster Classification and Victim Detection using Multi-Task Learning

Abstract

Recent advances in deep learning and computer vision have transformed surveillance into an important application for smart disaster monitoring systems. Based on the detected number of victims and activity of disasters, emergency response unit can dispatch manpower more efficiently, which could save more lives. However, most of existing disaster detection methods fall into the class of single-task learning, which can either detect victim or classify disaster. In contrast, this paper proposes a YOLO-based multi-task model which performs the aforementioned tasks simultaneously. This is accomplished by attaching a disaster classification head model to the backbone of a victim detection model. The head model is inherited from the MobileNetv2 architecture, and we precisely select the backbone feature map layer to which the head model is attached. For the victim detection, results reveal that the solution achieves up to 0.6938 and 20.31 in terms of average precision and frame per second, respectively. Whereas for the disaster classification, the algorithm is comparable with most deep learning models that are specifically trained for single task. This shows that our solution is flexible and robust enough to handle both victim detection and disaster classification.