Robot-assisted natural disaster management is recently employed to aid human rescuers at diverse disaster sites. Due to its compactness and availability, drone has become an effective tool for searching survivors from confined space such as collapsed building or underground area. However, the current scope of research in this field is limited because the research tends to focus on increasing accuracy of 3d mapping, constructed by controlling quadrotor flight at disaster sites. Perceiving disaster environment is necessary for rescue mission, but finding victims at the earliest time is more critical for practical rescue operations. In this work, we propose an overall architecture for drone hardware that enables fast exploration of GPS-denied environment, and practical methods for victim detection are introduced. We employ DJI Matrice 100 and utilize hokuyo lidar for global mapping and Intel RealSense for local mapping. Our results show that fusing these sensors can assist rescuers to find victims of natural disaster in unknown environments, and the detection system is insensitive to illumination change.