Annu Int Conf IEEE Eng Med Biol Soc. 2021 Nov;2021:5067-5071. doi: 10.1109/EMBC46164.2021.9629840.
The elderly fall detection is one critical function in health of the elderly. A real-time fall detection for the elderly has been a significant healthcare issue. The traditional video analysis on cloud has large communication overhead. In this paper, a fast fall detection system based on the spatio-temporal optical flow model is proposed, which is further deeply compressed by a structured tensorization towards an implementation on edge devices. Firstly, an object extractor is built to extract motion objects from video clips. The spatio-temporal optical flow model is formed to estimate optical flow fields of motion objects. It can extract features from objects and their corresponding optical flow fields. Then these two features are fused to form new spatio-temporal features. Finally, the tensor-compressed model processes the fused features to determine fall detection, where the strongest optical field would indicate the fall. We conduct experiments with Multicam and URFD datasets.Clinical relevance- It demonstrates that the proposed model achieves the accuracy of 96.23% and 99.37%, respectively. Besides, it attains the inference speed of 83.3 FPS and storage reduction of 210.9×. Our work is further implemented on an AI acceleration core based edge device, and the runtime is reduced by 9.21×.This high performance system can be applied to the field of clinical monitoring in the future.
老年人跌倒检测是老年人健康的一个关键功能。实时的老年人跌倒检测一直是一个重要的医疗保健问题。传统的基于云的视频分析具有较大的通信开销。在本文中，提出了一种基于时空光流模型的快速跌倒检测系统，该系统通过结构化张量化进一步深度压缩，以便在边缘设备上实现。首先，构建了一个目标提取器，从视频片段中提取运动目标。形成时空光流模型来估计运动目标的光流场。它可以从对象及其对应的光流场中提取特征。然后将这两个特征融合形成新的时空特征。最后，张量压缩模型处理融合特征以确定跌倒检测，其中最强的光场表示跌倒。我们在 Multicam 和 URFD 数据集上进行了实验。临床相关性-该模型在 Multicam 和 URFD 数据集上分别实现了 96.23%和 99.37%的准确率。此外，它还实现了 83.3 FPS 的推理速度和 210.9×的存储减少。我们的工作进一步在基于 AI 加速核的边缘设备上实现，运行时间减少了 9.21×。这个高性能系统未来可以应用于临床监测领域。
