.png){kind=icon}
.png){kind=icon}
Return
Motion and torso-guided frame distillation for optimized learning-based fall detection
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-026-04447-5.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Timely and accurate fall detection remains a critical challenge in healthcare, particularly for older adults living independently, as falls can cause severe injuries or fatalities. Although vision-based deep learning methods show strong potential, their high computational demands and privacy concerns limit real-world deployment. Advances in edge computing provide an opportunity for privacy-preserving, on-device video analysis, enabling real-time fall detection without reliance on cloud infrastructure. However, edge-based approaches must maintain a careful balance between computational efficiency, detection accuracy, and low latency. To address these requirements, we propose an optimized lightweight hybrid architecture that combines a data distillation-based frame selection strategy with a compact convolutional neural network (CNN) for efficient spatial feature extraction and a Bidirectional Long Short-Term Memory (BiLSTM) network for temporal dynamics modeling. The frame distillation mechanism condenses video content by leveraging motion cues and torso displacement, minimizing redundancy while preserving the most informative frames. Extensive experiments on the URFD, Le2i Fall Detection, and MCF datasets, under both intra-subject and cross-dataset evaluation protocols, demonstrate that the proposed framework achieves robust fall detection performance, with promising cross-dataset results in several settings, outperforming state-of-the-art methods and supporting low-latency, privacy-aware deployment on resource-constrained edge devices.
Keywords:
Fall detection
Data distillation
Computer vision
Depthwise separable CNN
BiLSTM