本文已被：浏览 6471次 下载 0次
中文摘要: 跌倒是导致老年人受伤甚至死亡的主要原因。准确及时的跌倒检测系统可以帮助跌倒者获得紧急救援。 目前基于传感器的跌倒检测方法主要利用人工设计提取的信号特征来区分跌倒和非跌倒运动，但人工提取的特征往往会限制算法的精确度，增大算法时延。为提高跌倒检测的精确度和实时性，本文提出了一种基于深度学习的跌倒检测算法。该算法可以自动提取数据特征，实现从原始数据到检测结果的端到端的处理。算法模型主要由两层级联的长短期记忆（Long Short-Term Memory, LSTM）循环神经网络组成，通过神经网络提取加速度计和陀螺仪数据内部的特征，并判断是否有跌倒状况发生。我们使用两个公开数据集MobiAct和SisFall对算法性能进行评估。 实验结果显示，算法在两个数据集都达到了较高的精确度（99.58%以上）和较低的时延（2.2毫秒以内）。
Abstract:Falls are a major cause of injury or death to the elderly. A timely and accurate fall detection system can help the fallen people get emergency rescue. Current sensor-based fall detection methods rely on manually extracted features to distinguish falls from activities of daily living (ADLs), which may limit the accuracy or latency of the algorithms. In this paper we propose a deep learning model called DeepFall for fall detection with automatic feature extraction. DeepFall is a stacked Long Short-Term Memory (LSTM) neural network that can directly analyse raw sensor data and judge whether a fall occurred. It is an end-to-end model and requires little pre-processing. The model was evaluated with two datasets: MobiAct and SisFall. It achieved high accuracies (over 99.58%) and low latencies (within 2.2 milliseconds) with both two datasets.
文章编号：cg18000453 中图分类号： 文献标志码：
基金项目:国家自然科学基金项目
引用文本：