Baseline Models for Action Recognition of Unscripted Casualty Care Dataset

This paper presents a comprehensive framework of datasets and algorithms for action recognition in scenarios where data is scarce, unstructured, and unscripted. The long-term objective of this work is an intelligent assistant to the medic, a surrogate buddy, that can tell the medic what needs to get done in every step of trauma resuscitation. As an essential part of this objective, we collected datasets and developed algorithms suitable for emergent contexts, such as casualty care in the field, disaster response and recovery scenarios, and other related high-risks/high-stakes scenarios where real-time decision-making is crucial. The proposed framework enables the development of new algorithms by providing a standardized set of evaluation metrics and test cases for assessing their performance. Ultimately, this research seeks to enhance the capabilities of practitioners and emergency responders by enabling them to better anticipate and recognize actions in challenging and unpredictable situations. Our dataset, referred to as Trauma Thompson, includes Tourniquet Application, Tracheostomy, Tube Thoracostomy, Needle Thoracostomy, and Interosseous Insertion procedures. The proposed algorithms based on the relative position embedding for the Vision Transformer referred as to ReVit, can achieve competitive performance with the state-of-art algorithms on our dataset.