Victim detection and localisation in an urban disaster site

In this study, we model the disaster victim detection problem as a sub-problem of a larger casualty assessment problem, and propose a framework to solve it. The framework of algorithm independent components contains a victim detector, detection history component and a human robot interaction component that presents information obtained by the robot in a meaningful manner. The algorithm independence of the victim detector component is demonstrated by experiments conducted in a simulated disaster scenario using a simple body parts detector that uses HOG features with an SVM classifier, and the state-of-the-art DPM body parts detector. A FastSLAM based mapping component is used to keep track of unique detections and the information is presented via a tab based user interface. The experiments demonstrate the effectiveness of the framework components with the rescue robot correctly identifying the victims and presenting a map of the disaster location with victim markers.