Parlak Polatkan, Siddika. Object detection and activity recognition in dynamic medical settings using RFID. Retrieved from https://doi.org/doi:10.7282/T3MP5209
DescriptionEstablishing context-awareness is key to develop automated decision support systems for dynamic and high-risk scenarios, where a critical component of context is the current activity. This thesis addresses the RFID-based detection of used medical objects with the ultimate goal of recognizing medical activities. We set trauma resuscitation, the initial treatment of a severely injured patient in the emergency department, as our target domain. We first describe the process of introducing RFID technology in the trauma bay. We analyzed trauma resuscitation tasks, photographs of medical tools, and videos of simulated resuscitations to gain insight into resuscitation tasks, work practices and procedures, as well as the characteristics of medical tools. Next, we propose and evaluate strategies for placing RFID tags on medical objects and for placing antennas in the environment for optimal tracking and object detection. We also discuss implications for and challenges to introducing RFID technology in other similar settings characterized by dynamic and collocated collaboration. Next we evaluate the use of RFID technology for object detection and activity recognition in a realistic setting. We tagged 81 medical objects and eight providers in a real trauma bay and recorded RFID signal strength during 32 simulated resuscitations. We extracted descriptive features and applied machine-learning techniques to monitor object use. We achieved accuracy rates of >90% when identifying the instance of a particular object type that was used during a resuscitation. Performance for detecting the usage interval of an object depended on various factors specific to the object. Our results also provide insights into the limitations of passive RFID and areas in which RFID needs to be complemented with other sensing technologies. We also investigated the usability of object motion and location cues for activity recognition by conducting motion detection and localization experiments under challenging scenarios. Using statistical methods, we were able to detect object motion with an accuracy of 80%, and predict the zone where the object is located with an accuracy of 86%.