Advancing Human Fall Detection with TsetFall: A Comprehensive, Fine-Grained, and Publicly Available Dataset

Human fall detection research remains an under-developed area, particularly with respect to utilizing dynamic images for analysis. Existing fall detection datasets often suffer from size, diversity, and representativeness limitations, as they generally comprise a small number of videos featuring a limited range of actions, camera perspectives, and lighting conditions. Moreover, these datasets typically lack the complexity of real-world fall scenarios, as they often exclude distracting objects and offer only one or two camera angles. To address these challenges, this work introduces TsetFall, a comprehensive, fine-grained, and publicly available dataset generated for fall detection research. Employing a novel AI technique devised by the present authors, the dataset annotation process was notably expedited. By bridging the current gaps in available resources, the TsetFall dataset serves as a valuable benchmark for advancing the field and tackling the multifaceted nature of human fall detection.