Infrared Image Human Detection for Search and Rescue Processes (SARHuman)

The environments we used for simulating post-disaster ruins were cluttered and rocky. Most of our images were shot at night to simulate the poorly lit interior of the ruins. Each image has 1-3 bodies, which are in different poses. The photographed object is obscured by debris, such as rocks, to simulate what actually happens in a ruined environment. We also photographed the objects from far and near, ensuring that there would be human objects at different scales in the dataset. In addition, we consider that a SAR robot that can search deep into the ruins will not be too large and can carry an infrared camera with limited resolution, so we use a FLIR infrared camera with a resolution of 320×320 to capture images.
We removed blurred images. We continuously adjusted the captured images so that the images in each background were evenly distributed. We obtained a dataset containing 5189 infrared images named SARHuman. Here are some examples of the dataset.