After a disaster such as an earthquake, if evacuation route information of evacuees to the outside of a building can be recorded on their smartphones on the principle of PDR(Pedestrian Dead Reckoning) and passed to rescue staffs, it would be useful for them to reach the location of injured persons in the building. Assuming this scenario, we have proposed the motion recognition scheme of a person using a smartphone's rotation vector sensor, an acceleration sensor and so on. The limitation of this scheme was that it needs sensor data of the person obtained in the walking state during an evacuation as reference data. In this paper, we propose a method to acquire reference data of the walking state regarding the target person in his or her daily life in advance and use it to recognize the moving state at the time of evacuation. Experimental results show that recall rates were over 80% in each motion state of walking, running, descending stairs, and ascending stairs.