Human Detection and Localization in Indoor Disaster Environments Using UAVs

This paper presents human detection and localization module for GPS-denied indoor environments in disaster scenario. We evaluate the performance of three human detection algorithms: Histogram of Oriented Gradients, YOLO v3, and Mask RCNN using occluded human dataset. We compare them on performance metrics such as runtime, precision, and recall. YOLO v3 has the best trade-off between accuracy and computational runtime. This model is used to implement a human detection and localization module using ROS. The module was tested in simulated Gazebo environments that depict indoor disaster scenarios. The localization errors of detected humans show the this approach is suitable for localizing humans which can help in better planning of rescue operations.