HUST-FALL

HUST-FALL is a fine-grained text-video multimodal dataset specifically designed for unconstrained fall detection in the wild. It integrates videos from diverse sources\u2014including open-source datasets and online platforms\u2014covering a wide range of fall scenarios, subjects, environments, lighting conditions, and occlusion patterns. Unlike existing fall detection datasets that are primarily collected under controlled indoor settings, HUST-FALL captures the complexity and variability of real-world falls. Each video is further annotated with structured fine-grained textual descriptions across four key dimensions: body dynamics, environmental interaction, subject attributes, and scene context. These rich annotations serve as high-level semantic guidance to facilitate vision-language reasoning, making HUST-FALL a challenging yet valuable benchmark for advancing generalizable fall detection models in real-world applications.