Deep Learning for Multi-person Detection and Tracking in Mass Casualty Incidents

Abstract: To evaluate and prioritize the injured during an incident of mass casualty, obtaining situational and positional awareness of the site is essential. There are situations where the first responders (nurses, firemen, police, etc.) cannot gain this perception by themselves. This perception can be gained from a distance by utilizing an unmanned aerial vehicle (UAV) equipped with a camera together with a deep learning (DL) system. This thesis evaluates several robust deep learning methods to perform multi-object detection and tracking with a moving RGB and thermal infrared (TIR) camera. The TIR camera becomes important during nightly conditions and when there is a need to detect and track the injured through smoke and debris. The methods evaluated for detection are YOLOv5 and Faster R-CNN, while SORT and Deep SORT are evaluated for tracking. The detection results show that YOLOv5 performs better than Faster R-CNN in the RGB domain, while Faster R-CNN outperforms YOLOv5 in the TIR domain. Overall, the detection performance is better in the RGB domain compared to the TIR domain when considering the specific datasets evaluated in this thesis. The tracking results show that Deep SORT outperforms SORT, where the performance increase is more significant in the RGB domain compared to the TIR domain. Similar to the detection task, the tracking performance proves in general to be better in the RGB domain compared to the TIR domain.