Disaster Information Acquisition And Map Building For MAVs In The Internal Scene Of A Building

Before rescuers enter the disaster scene,it is a good choice for rescuers to utilize robots to evaluate disaster scenarios and construct corresponding environmental maps,which can help disaster rescuers to accomplish the rescue task accurately and effectively.The micro air vehicles(MAVs)has good maneuverability and can fly into high buildings through damaged windows,which is more practical than ground mobile robots to enter buildings after disasters.In this paper,we studied and proposed solutions to the problem of disaster information acquisition and map building for MAVs in the internal scene of a building.In order to solve the problem of a MAV entering building through broken window,we propose a window detection algorithm based on keypoints detection.Different from common indoor objects,the center of the broken window area for a MAV's passing through is usually with irregular background information and it cannot be solved with common object detection algorithm.To avoid the interference of the uncertain background information,keypoint heatmaps are adopted to regress the corner position of the window,so that it pays more attention to the edge information.The proposed algorithm has a good performance in both data set and real-world testing.While a MAV entering the building through the window,it needs to use infrared vision and visible vision sensors to detect injured person and other indoor objects simultaneously.Compared with classical single-stage object detection algorithms,YOLOv2 is selected as the benchmark algorithm due to its high performance in both speed and accuracy,which can meet the requirements of limited computing resources of MAVs.YOLOv2 uses the low-resolution feature maps to detect objects,which leads to the problem of inaccurate object center location in our tests.For solving this problem,we propose a method to upsample the feature map of backbone network 4 times,and use the central heatmap to supervise the network training.By using the high-resolution feature map,the accuracy of YOLOv2 algorithm in the wounded detection and indoor object detection can be improved effectively.Based on the above detection results,we present a solution to accomplish semantic map building in an indoor environment.Considering that the map building by using the ORBSLAM2 can only obtain geometric information but not semantic information,we use the 2D object detection algorithm to detect the key-frame of ORB-SLAM firstly,and then project the detected 2D objects in the key-frame into the 3D map.At the same time,the problem of data association between different frames is also solved,so that the MAV can build a map with semantic labels.We have tested the proposed object detection and semantic map building method in OpenLORIS-Scene dataset,and the experimental results show that our method is valid.