Multimedia Emergency Event Extraction and Modeling Based on Object Detection and Bi-LSTM network

Abstract

Worldwide public emergency events are becoming a significant problem that is threatening the peace and development of the world. Facing at the frequent occurrence of violent and emergency cases around the world, there is no country or region can be entirely outside the incident. Typical events including Hong Kong occupy central incident and Coronavirus have brought significant impact on countries and regions all around the world. Inspired by the occurrence of such events, we proposed two general models that can detect the objects in the events and find the pattern of the events for further prediction and classification.The current object detection and classification models mostly focus on one specific type of objects. When dealing with universal model, most of them will have inferior performance in single model due to the construction of universal model. In this paper, we proposed a new general object detection model to address to the problem that outperforms most detection models that focus on single object detection. In the first part, we proposed a web crawler to find the related emergency events videos for training. Then, we extract the key frames of these videos for the effective training progress. The Universal Event Extraction Model we proposed includes large amounts of objects including human gathering, geological location, police siren, protest and explosion scenarios and the arguments of the scenarios. In the last part of the paper, we applied Graph Attention Model+Bi-LSTM to find the pattern of the emergency events for further prediction and classification.