3DKD - An Effective Knowledge Distillation-Based Model for Human Fall Detection

Abstract

Accidental falls are a commonly encountered occurrence in various groups of people, encompassing children, the elderly, and adults alike. The prompt detection of human falls represents a paramount method for mitigating the substantial risks associated with loss of self-control, fatality, or physical harm. Moreover, such proactive measures hold the potential to curtail healthcare expenses. Consequently, there exists a pressing need for research and development of systems aimed at detecting and facilitating the rescue of individuals involved in falls. In this regard, the proposed model, namely 3DKD, is constructed by employing the technique of knowledge distillation in conjunction with the attention mechanism, thereby yielding significant improvements in computer vision tasks within the same model layer. This innovative approach is anticipated to contribute to the creation of more precise deployment models capable of operating efficiently on weak devices.