Many people have faced mortal risks due to sudden disasters such as earthquakes, fires, and terrorisms, etc. In disasters where most people become panic, it is important to grasp disaster positions immediately and to find out some appropriate evacuation routes. We previously proposed the specific evacuation support system named as Emergency Rescue Evacuation Support System (ERESS). ERESS is based on Mobile Ad-hoc network (MANET) and aims to reduce the number of victims in panic-type disasters. This system consists of mobile terminals with advanced disaster recognition algorithm and various sensors such as acceleration, angular velocity and earth magnetism. However, the former ERESS did not have the clear criteria to detect the disaster outbreak. In this paper, we propose a new disaster recognition algorithm by Support Vector Machine (SVM) which is a kind of machine learning. In this method, an ERESS mobile terminal learns the behaviors of its holder by SVM. The SVM acquires the decision boundary based on the sensing data of the terminal holder, and it is judged whether to be the emergency. We show the validity of the proposed method by panic-type experiments.