Conference Agenda

Traditional fall detection systems are mostly limited to binary classification, failing to describe event dynamics or severity. Large Multimodal Models (LMMs) can perform detailed semantic and biomechanical analysis from video, but their high computational demands limit continuous real-time deployment. This study presents a proof-of-concept hybrid architecture fusing a low-power wearable sensor trigger with a two-stage cloud-based LMM video analysis pipeline to characterize falls and reduce false alarms. Upon wearable detection of a candidate event, a continuously recording camera saves a synchronized video clip. A first LMM stage pre-screens the clip to filter false positives; if a fall is confirmed, a second LMM stage extracts structured fall-related parameters (initiating cause, fall direction, protective hand response, and impacted body parts) to support caregiver alerting. Validation of the pre-screening stage on 79 clips spanning three simulated fall types and three confounding activity classes yielded 87.3\% overall accuracy (precision $= 80.6\%$, recall $= 90.6\%$, specificity $= 85.1\%$). In Stage~2, evaluated on 28 real-world surveillance videos, the LMM achieved 96.4\% accuracy for initiating cause, 85.7\% for protective hand response, 71.4\% for fall direction, and a mean Jaccard index of 0.685 for impacted body-part identification. The results demonstrate the potential of integrating wearable fall detection with LMMs for structured fall characterization, although residual misclassification errors require further investigation.