Data Augmentation Method for Improving Object Detection Accuracy of Recumbent Human in Disaster

For this study, we specifically examine object detection methods for finding victims at a facility who are in a lying posture because of injury or for other reasons. Most large-scale human training data available on the internet are based on the standing postures of people. Lying postures are rare. Furthermore, obtaining large numbers of images of people lying down in disaster situations and creating training data are extremely difficult. Therefore, it is necessary to learn object detection models that handle diverse postures. For this study, we assess data augmentation. Because the image of a person in a reclined posture is regarded as one in which the head and body are aligned side by side, data extension of 90° rotation to the existing image of a person in a standing posture can be used to enhance the image of a person in a reclining posture. For this study, the object detection algorithm SSD is augmented with rotational data augmentation to enhance the detection of lying human figures. However, it is unknown whether the rotation data enhancement is effective for detecting lying persons or not. For an evaluation experiment, we changed the rotation rate arbitrarily and compared the models learned with each rate. The data expansion of the rotation was evaluated using the detection accuracy for normal and recumbent human figures. We confirmed that the detection accuracy of the lying person improved at a rotation of 25%.