WO2021212883A1 - 一种基于智能移动终端的跌倒检测方法 - Google Patents

一种基于智能移动终端的跌倒检测方法 Download PDF

Info

Publication number
WO2021212883A1
WO2021212883A1 PCT/CN2020/137260 CN2020137260W WO2021212883A1 WO 2021212883 A1 WO2021212883 A1 WO 2021212883A1 CN 2020137260 W CN2020137260 W CN 2020137260W WO 2021212883 A1 WO2021212883 A1 WO 2021212883A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
fallnet
alarm
data
fall detection
Prior art date
Application number
PCT/CN2020/137260
Other languages
English (en)
French (fr)
Inventor
邢建川
谭玉博
王翔
刘懿尧
王雨轩
陈思芹
王宇萌
苑舒雨
Original Assignee
电子科技大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 电子科技大学 filed Critical 电子科技大学
Publication of WO2021212883A1 publication Critical patent/WO2021212883A1/zh

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/103—Detecting, measuring or recording devices for testing the shape, pattern, colour, size or movement of the body or parts thereof, for diagnostic purposes
    • A61B5/11—Measuring movement of the entire body or parts thereof, e.g. head or hand tremor, mobility of a limb
    • A61B5/1116—Determining posture transitions
    • A61B5/1117—Fall detection
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/74—Details of notification to user or communication with user or patient ; user input means
    • A61B5/746—Alarms related to a physiological condition, e.g. details of setting alarm thresholds or avoiding false alarms
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/74—Details of notification to user or communication with user or patient ; user input means
    • A61B5/7465—Arrangements for interactive communication between patient and care services, e.g. by using a telephone network
    • A61B5/747—Arrangements for interactive communication between patient and care services, e.g. by using a telephone network in case of emergency, i.e. alerting emergency services
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G06F18/24137—Distances to cluster centroïds
    • G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B2503/00—Evaluating a particular growth phase or type of persons or animals
    • A61B2503/08—Elderly

Definitions

  • the invention relates to the technical field of intelligent detection, in particular to a fall detection method based on an intelligent mobile terminal.
  • the present invention proposes a fall detection method based on a smart mobile terminal.
  • the fall detection method designs a FallNet model. With a small increase in parameters, the 17-category classification effect of the FallNet model reaches 98.59%, and the two-category ROC The Area Under Curve (AUC) value enclosed by the curve and the coordinate axis is increased to 0.9984.
  • the APP can recognize human activities, and can also issue alarms and alarms for human falls, which can realize intelligent monitoring of the health status of the elderly.
  • a fall detection method based on a smart mobile terminal includes the following steps:
  • Step 1 Collect well-represented human activity data in terms of height, weight, and age, and construct a data set, and then preprocess the data set, and use 80% of the human activity data in the preprocessed data set as the training set , Using 20% of human activity data as the test set;
  • Good representation means that the physical data of these people can reflect the overall physical data of human beings to a certain extent.
  • Table 1 is a schematic table of the characteristics of the subjects, as shown in Table 1 below.
  • Step 2 Use feature engineering technology to perform feature extraction and analysis on the data set, use principal components analysis (PCA) dimensionality reduction technology to analyze feature vectors, and select high-quality features for the next step of training;
  • PCA principal components analysis
  • High-quality features refer to features that use PCA dimensionality reduction technology to reduce the dimensionality of their feature vectors, while reducing the indicators that need to be analyzed, the original indicators contain less loss of information, and the collected data can still be fully analyzed.
  • Step 3 Based on the Long Short Term Memory-Fully onvolutional Network (LSTM-FCN) model, and then design an improved FallNet model, in the Fully Convolutional Neural Network (FCN) network Add the Batch Normalization layer before the Long Short Term Memory (LSTM) network.
  • the Batch Normalization layer normalizes the input data, and the normalized data is input to the full volume machine module and the LSTM module, and Add the global maximum pooling and global average pooling layers to the input layer to extract the amplitude characteristics of the input sequence, and finally perform the same operation after each corresponding convolution activation module, and then use the training set to perform the FallNet model Train and use the test set to test the FallNet model;
  • the input sequence refers to the acceleration data of human activities; the input sequence of step three is a sequence composed of the high-quality features of step two.
  • the amplitude features of high-quality features can still contain a large amount of original data information after dimensionality reduction.
  • the same operation refers to adding the global maximum pooling and global average pooling layers after each convolution activation module.
  • Step 4 Design the APP.
  • the APP adopts a short-term-long-term continuous monitoring method, builds the trained FallNet model in the mobile device, and then performs sliding window processing on the collected human activity data, and then performs the sliding window processing on the collected data according to the FallNet model Human activity data is used for fall detection, and a local alarm module and a remote alarm module are set in the mobile device, and the local alarm module and remote alarm module are used to alarm the results of the fall detection for help.
  • a further improvement lies in that: when the data set is preprocessed in the first step, the data set is randomly shuffled by means of data division.
  • the LSTM module is a cyclic neural network layer containing 8 LSTM units, and the LSTM module extracts features according to the input time series; according to the input acceleration data, extracts 8 feature values and combines them with all levels of the FCN network Features are integrated for use in the output layer.
  • the time series here is the series entered in step three.
  • the high-quality features determine the behavior that needs to be studied; the acceleration data is the acceleration data of the behavior to be studied measured by the mobile phone; the acceleration data of the behavior is sequential and constitutes a time series.
  • a further improvement is that: a parameter loss layer is connected behind the LSTM module.
  • the parameter loss layer randomly loses some features, so as to retrain the network and prevent overfitting.
  • the partial features are features that randomly lose a set number of features.
  • a further improvement is that when the APP adopts a short-time-long-time continuous monitoring method in the step four, it first performs real-time action detection in a short period of time, and then performs long-term human body state monitoring.
  • Short-term and long-term refer to the time interval for monitoring the state of the human body; short-term detection is real-time motion detection. Long time refers to the monitoring of human body status over a long period of time, which can be regarded as a historical record.
  • a further improvement is that: when the human body is monitored for a long period of time in the fourth step, the action monitoring sequence is used as the data judgment basis.
  • the local alarm module automatically starts the pre-alarm program, and when the user continues to be detected at the specified time Normal activity, cancel the alarm, otherwise start the automatic alarm procedure.
  • the user's actions are judged based on the above-mentioned time sequence, and the action detection sequence is obtained.
  • the specified time can be within 10 seconds, if the user continues to be detected as normal activities, the system cancels the alarm.
  • a further improvement is that the process for the local alarm module to automatically start the pre-alarm program is: when it detects that the user has fallen, the mobile device sends out an alarm voice to ask for local help; at the same time, it sends out a voice prompt and asks within a preset waiting time Whether the user’s fall is true and whether the alarm needs to be canceled. If the user chooses to cancel, the alarm will be canceled. Otherwise, the alarm will be immediately alarmed at the end of the specified time; when no user action is detected, the same voice inquiry will be performed. When the user presses the cancel button, then Cancel the alarm, otherwise it is considered that the user has fallen and is in a dangerous situation and needs to contact the guardian and call the emergency number immediately.
  • a further improvement is that the remote alarm module is used for when the user is confirmed to have fallen, while the local alarm module alarms for help, the remote alarm module synchronously sends the current fall information and location to the guardian's mobile phone and dials an emergency call.
  • the beneficial effects of the present invention are: through the design of the FallNet model, the 17-category classification effect of the FallNet model reaches 98.59%, and the two-category AUC value is increased to 0.9984 by designing the FallNet model with a small number of parameters.
  • the fall detection APP it can identify human activities, and can also issue alarms and alarms for human falls, which can realize intelligent monitoring of the health status of the elderly, and the monitoring process is highly real-time.
  • FIG. 1 Schematic diagram of the data distribution of each category of the data set of the present invention
  • Figure 2 is a schematic diagram of action recognition and fall detection of the present invention
  • Figure 3 is a schematic diagram of the network structure of the FallNet model of the present invention.
  • this embodiment proposes to take the open-source comprehensive data set UniMiB-SHAR as an example.
  • the data set is randomly shuffled, 80% of the data is included in the training set, and the rest are used as the test set.
  • the distribution of category data is shown in Figure 1.
  • the abscissa represents the sample index of the data.
  • the broken line primary is the distribution of the original data set. After the data set is evenly shuffled, the curve shuffle is obtained.
  • the labels of each sample stage are uniformly random.
  • the experiment is based on the Windows 10 operating system, and the implementation language is python 3.6.5.
  • 100 rounds of training are performed on the training set, and the trained model weights are obtained, and then evaluated on the test set, and the experimental results are further studied and analyzed.
  • the UniMiB-SHAR data set is a smart phone-based human activity recognition data set of the University of Milan Bicocca
  • Five-fold cross-validation Divide the data set into multiple subsets uniformly and randomly. Taking five-fold cross-validation as an example, divide the data set into five subsets A, B, C, D, and E. In the training of a model, a total of Take 5 rounds of training, each time a different subset is used as the test set, and the remaining subsets are used as the training set. The specific process is shown in Table 2 below:
  • the parameter evaluation takes the mean value of five rounds of testing:
  • the model evaluation parameters are as follows:
  • the total recall is the average of the recall of each category:
  • the weighted average of recall and precision, F1-Score is also a balanced evaluation of accuracy and recall.
  • TP true
  • TN true negative
  • FP false positive
  • FN False negative
  • the prediction is negative, but the actual is positive.
  • the FallNet model of the present invention has advantages in various indicators: accuracy, recall, sensitivity and the like. In the evaluation indicators of 17 types of action classification, the accuracy rate reaches 98.59%, which is not only much higher than traditional machine learning methods, but also 0.42% higher than the LSTMFCN with an accuracy rate of 98.17. It can be used for actual human activity classification.
  • the FallNet model is a model that can be selected as a priority for human activity recognition.
  • Run test for the developed Android application The running test environment is MI8Lite, MIUI version 10.2, Android version 8.1.0, running memory 4.00GB, processor 8 cores up to 2.2GHz. The results show that the App can run normally and do fall detection work.
  • the 17-category classification effect of the FallNet model reached 98.59%, and the two-category AUC value was increased to 0.9984, and this model was used to design a fall detection APP, which can Human body activity can be recognized, and alarms and alarms can also be issued for human body falls, which can realize intelligent monitoring of the health status of the elderly, and the monitoring process is of high real-time.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mathematical Physics (AREA)
  • Veterinary Medicine (AREA)
  • Surgery (AREA)
  • Medical Informatics (AREA)
  • Animal Behavior & Ethology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Pathology (AREA)
  • Public Health (AREA)
  • Computing Systems (AREA)
  • Physiology (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Fuzzy Systems (AREA)
  • Signal Processing (AREA)
  • Psychiatry (AREA)
  • Business, Economics & Management (AREA)
  • Critical Care (AREA)
  • Emergency Management (AREA)
  • Emergency Medicine (AREA)
  • Nursing (AREA)
  • Dentistry (AREA)
  • Oral & Maxillofacial Surgery (AREA)

Abstract

一种基于智能移动终端的跌倒检测方法，包括以下步骤：采集在身高、体重、年龄方面有良好代表性的人体活动数据，并构建数据集；运用特征工程技术对数据集进行特征提取与分析，使用PCA降维技术对特征向量进行分析；基于LSTM-FCN模型设计FallNet模型，对FallNet模型进行训练；将训练好的FallNet模型内置在移动设备中进行跌倒检测；通过设计FallNet模型，在增加少量参数的情况下，FallNet模型的17类别分类效果达到了98.59％。能够对人体活动进行识别，对人体跌倒发出警报和报警，实现对老龄人群健康状态的实时监测。

Description

一种基于智能移动终端的跌倒检测方法
本申请要求于2020年04月20日提交中国专利局、申请号为202010309877.4、发明名称为“一种基于智能移动终端的跌倒检测方法”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及智能检测技术领域，特别是涉及一种基于智能移动终端的跌倒检测方法。
背景技术
随着国家开放二孩政策的施行，我国人口老龄化的趋势日趋明显，如何对于老龄人群健康状态进行智能监测已经成为一个重要的课题。人类活动识别(Human activity recognition,HAR)利用传感器数据来实时分辨活动，随着物联网的迅速发展，这种方法近年来受到了广泛的关注。对于身体衰弱的老人，撞击、跌倒这些影响因素都有可能造成不可挽回的伤害，如果能够设计出一种方便携带、灵敏度高、智能化的活动识别设备，无疑对于老人的健康监护具有重大意义，必将会给更多家庭带来福音。
近年来，对日常生活活动(ADLs)的识别与分类技术的研究有了长足的发展，通常是通过分析从传感器获得的信号对人类活动进行分类。跌倒检测精度比较高的方式是图像识别方法，但存在的问题是很难有条件在每个地方安装上摄像头，更不可能用这种方法对某一个人进行连续地检测；采用基于阈值的检测方法在进行跳跃或者缓慢跌倒的情况下则很容易发生误判、漏报，因此这种方法不具备广泛的评估能力。为了更好的对老龄人群健康状态进行智能监测，本发明提出一种基于智能移动终端的跌倒检测方法，以解决现有技术中的不足之处。
发明内容
针对上述问题，本发明提出一种基于智能移动终端的跌倒检测方法， 该跌倒检测方法通过设计FallNet模型，在增加少量参数的情况下，FallNet模型的17类别分类效果达到了98.59％，二分类ROC曲线下与坐标轴围成的面积(Area Under Curve，AUC)值增加到0.9984，APP能够对人体活动进行识别，也可以对人体跌倒发出警报和报警，可以实现对老龄人群健康状态的智能监测。
为实现本发明的目的，本发明通过以下技术方案实现：
一种基于智能移动终端的跌倒检测方法，包括以下步骤：
步骤一：采集在身高、体重、年龄方面有良好代表性的人体活动数据，并构建数据集，然后对数据集进行预处理，并将预处理后的数据集中80％的人体活动数据作为训练集，将20％的人体活动数据作为测试集；
良好代表性指可以这些人的身体数据在一定程度上可以反映人类总体的身体数据。
例如：身高：160至190厘米；体重：50至82公斤；年龄：18至60岁；表1为受试者特点示意表，如下表1所示。
表1
Figure PCTCN2020137260-appb-000001
步骤二：运用特征工程技术对数据集进行特征提取与分析，使用主成分分析(Principal components analysis，PCA)降维技术对特征向量进行分析，选取优质特征进行下一步训练；
优质特征指使用PCA降维技术对其特征向量进行降维之后，在减少需要分析的指标同时，原指标包含信息的损失较少，仍可对所收集数据进 行全面分析的特征。
步骤三：基于长短时记忆全卷积网络(Long Short Term Memory-Fully onvolutional Network，LSTM-FCN)模型，然后设计一种改进的FallNet模型，在全卷积神经网络(Fully Convolutional Network，FCN)网络和长短时记忆(Long Short Term Memory，LSTM)网络之前加上批量归一化Batch Normalization层，Batch Normalization层对输入数据进行归一化，归一化的数据输入全卷机模块和LSTM模块，并在输入层加上全局最大池化和全局平均池化层，用于提取输入序列的幅度特征，最后在相应的每一个卷积激活模块后面都进行相同的操作，然后利用训练集对FallNet模型进行训练，并利用测试集对FallNet模型进行测试；
输入序列指人体活动的加速度数据；步骤三的输入序列是步骤二的优质特征组成的序列。
优质特征的幅度特征在降维后仍能包含大量原始数据信息。
相同的操作指在每一个卷积激活模块后面都加上全局最大池化和全局平均池化层。
步骤四：设计APP，APP采取基于短时间-长时间连续监测的方式，将训练好的FallNet模型内置在移动设备中，然后对采集的人体活动数据进行滑动窗处理，然后根据FallNet模型对采集的人体活动数据进行跌倒检测，并在移动设备中设置本地报警模块和远程报警模块，利用本地报警模块和远程报警模块对跌倒检测结果进行报警求助。
进一步改进在于：所述步骤一中对数据集进行预处理时，利用数据划分手段将数据集随机打乱。
进一步改进在于：所述LSTM模块为包含8个LSTM单元的循环神经网络层，LSTM模块根据输入的时间序列提取特征；针对输入的加速度 数据，进行提取8个特征值，并与FCN网络的各级特征进行整合，用于输出层使用。
此处的时间序列为步骤三中的输入的序列。
优质特征确定了需要研究的行为；加速度数据是手机测得的要研究的行为的加速度数据；行为的加速度数据是有时间先后的，构成时间序列。
进一步改进在于：所述LSTM模块后面连接有参数丢失层，FallNet模型的每一轮训练中，参数丢失层随机丢失部分特征，实现重新训练网络、防止过拟合。
在训练时，让参数层某个神经元的激活值以一定的概率p停止工作；即，所有特征都有一定的概率进行丢失。因此，所述部分特征为随机丢失设定数量的特征。
进一步改进在于：所述步骤四中APP采取基于短时间-长时间连续监测的方式时，首先在短时间进行实时动作检测，然后进行长时间段的人体状态监测。
短时间和长时间指的是监测人体状态的时间间隔；短时间进行是实时动作检测。长时间是指长时间段的人体状态监测，可视为历史记录。
进一步改进在于：所述步骤四中进行长时间段的人体状态监测时，以动作监测序列作为数据判断依据，当发生跌倒动作，本地报警模块自动启动预报警程序，当在指定时间检测到用户继续正常活动，则取消报警，否则起动自动报警程序。
基于上述的时间序列判断用户的动作，得到动作检测序列。
在实际应用中，指定时间可以为10秒内，如果检测到用户继续正常活动，则***取消报警。
进一步改进在于：所述本地报警模块自动启动预报警程序的过程为：当检测到用户发生跌倒动作时，移动设备发出报警语音，进行本地求助；同时在预设置的等待时间内发出语音提示，询问用户跌倒是否属实，是否需要取消报警，如果用户选择取消，则取消报警，否则在指定时间结束时， 立即报警；当检测到用户无动作时，同样进行语音询问，当用户按下取消键，则取消报警，否则认为用户已经跌倒并且属于危险状况，需要立即联系监护人和拨打急救电话。
进一步改进在于：所述远程报警模块用于当确认用户跌倒时，在本地报警模块报警求助的同时，远程报警模块同步将当前跌倒信息与位置发送到监护人手机并且拨打急救电话。
本发明的有益效果为：本发明通过设计FallNet模型，在增加少量参数的情况下，FallNet模型的17类别分类效果达到了98.59％，二分类AUC值增加到0.9984，并且应用这一模型，设计出了跌倒检测APP，能够对人体活动进行识别，也可以对人体跌倒发出警报和报警，可以实现对老龄人群健康状态的智能监测，且监测过程实时性高。
说明书附图
下面结合附图对本发明作进一步说明：
图1本发明数据集各类别数据分布示意图；
图2为本发明动作识别与跌倒检测示意图；
图3为本发明FallNet模型网络结构示意图。
具体实施方式
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
根据图1、2、3所示，本实施例提出以开源的综合数据集UniMiB-SHAR为例，把数据集随机打乱，80％的数据归入训练集，其余作测试集，数据集各类别数据分布如图1所示，图1中横坐标代表数据的样本索引，从0到11770一共11771个样本数据，纵坐标代表各样本数据 的标签，从0到16共17种分布可能。折线primary是原始数据集的分布情况，对数据集均匀打乱后，得到曲线shuffle，各个样本阶段的标签是均匀随机的。
在UniMiB-SHAR数据集上统一训练和把训练好的模型进行五折交叉验证。实验基于Windows10操作***，实现语言为python3.6.5。每次验证都在训练集上训练100轮，获得训练好的模型权重，然后在测试集上进行评估，并对实验结果做进一步研究分析。UniMiB-SHAR数据集为米兰比科卡大学基于智能手机的人类活动识别数据集
五折交叉验证：把数据集均匀随机划分成多个子集，以五折交叉验证为例，把数据集划分为A、B、C、D、E五个子集，在一个模型的训练中，一共采取5轮训练，每次训练分别取不同的一个子集作为测试集，其余子集共同作为训练集，具体过程如下表2所示：
表2
训练轮数 训练集 测试集 测试参数
1 ABCD E a1
2 ABCE D a2
3 ABDE C a3
4 ACDE B a4
5 BCDE A a5
参数评估取五轮测试的均值：
a＝(a1+a2+a3+a4+a5)/5
模型评估参数如下：
精度：
测量被归为阳性的样本中真实样本的数量，总精度是每类精度的平均值：
Precision＝TP/(TP+FP)
召回率(敏感度)：
测量一个类别的总样本中正确分类的样本数，总召回是每一类召回的平均值：
准确度：
Recall＝TP/(TP+FN)
测量正确预测的标签与所有预测的比例：
Accuracy＝(TP+TN)/(TP+TN+FP+FN)
F1-score：
召回率和精度的加权平均，F1-Score也是对准确率与召回率的一个均衡评价。
F1-score＝2TP/(2TP+FP+FN)
其中，Truepositives(TP，真正)：预测为正，实际为正；True negatives(TN，真负)：预测为负，实际为负；Falsepositives(FP，假正)：预测为正，实际为负；False negatives(FN，假负)：预测为负，实际为正。
然后测试主流的各种机器学习算法在二分类任务下的表现，使用的评估参数有召回率(recall)、精确度(precision)、准确率(accuracy)、F1-Score(f1)和ROC曲线面积(AUC)，得到表3的结果：
表3
算法 recall precision accuracy f1 auc acc17
knn1_uniform 0.95804196 0.99878493 0.98428875 0.97798929 0.97868698 0.85987261
knn1_distance 0.95804196 0.99878493 0.98428875 0.97798929 0.97868698 0.85987261
GaussianNB 0.93123543 0.94110719 0.95371550 0.93614528 0.94891765 0.58980892
BernoulliNB 0.67599068 0.46068308 0.59363057 0.54794521 0.61120843 0.15966030
svm_sigmoid 0.65967366 0.93245470 0.85859873 0.77269625 0.81614278 0.29681529
svm_linear 0.97785548 0.98705882 0.98726115 0.98243560 0.98525372 0.74734607
svm_poly 0.00000000 0.00000000 0.63566879 0.00000000 0.50000000 0.16687898
svm_rbf 0.93822844 0.88461539 0.93290871 0.91063348 0.93404408 0.46963907
tree_c45 0.96853147 0.96627907 0.97622081 0.96740396 0.97457970 0.76687898
tree_bag 0.98951049 0.98720930 0.99150743 0.98835856 0.99108123 0.89426752
tree_cart 0.98135198 0.97793264 0.98513800 0.97963933 0.98432997 0.77791932
tree_forest 0.46386946 1.00000000 0.80467091 0.63375796 0.73193473 0.61019108
xgboost 0.99067599 0.99067599 0.99320595 0.99067599 0.99266599 0.87600849
ConvNet 0.98203600 0.99153600 0.99065800 0.98676300 0.98871500 0.85562600
HyBrid 0.98203600 0.97735400 0.98556300 0.97698900 0.98476800 0.87855600
LstmFCN 0.99640700 1.00000000 0.99872600 0.99820000 0.99820400 0.98174100
FallNet 0.99760500 0.99880100 0.99872600 0.99820300 0.99847300 0.98598700
从表3可以看出：其中FallNet最优，LSTMFCN次之，Hybrid和ConvNet再次之。LSTMFCN和FallNet在17类别分类中和2分类的其他评价指标中都达到很好的效果，各方面参数超过98％，而Hybrid和ConvNet在2分类的指标中得分较高，但是在17类别分类中要落后于FallNet和LSTMFCN很多。
本发明的FallNet模型在各项指标：准确率、召回率、灵敏性等方面都占据优势。在17类动作分类的评价指标中，准确率达到98.59％，不仅比传统的机器学习方法高很多，而且比准确率达到了98.17的LSTMFCN提升 了0.42个百分点，可以用于实际的人体活动分类，FallNet模型是一种可以优先选择作为人体活动识别的模型。
对App运行测试：
针对开发完成的安卓应用进行运行测试，运行测试环境为MI8Lite，MIUI版本10.2，安卓版本8.1.0，运行内存4.00GB，处理器8核最高2.2GHz。结果表明：App可以正常运行并做跌倒检测工作。
本发明通过设计FallNet模型，在增加少量参数的情况下，FallNet模型的17类别分类效果达到了98.59％，二分类AUC值增加到0.9984，并且应用这一模型，设计出了跌倒检测APP，能够对人体活动进行识别，也可以对人体跌倒发出警报和报警，可以实现对老龄人群健康状态的智能监测，且监测过程实时性高。
以上显示和描述了本发明的基本原理、主要特征和优点。本行业的技术人员应该了解，本发明不受上述实施例的限制，上述实施例和说明书中描述的只是说明本发明的原理，在不脱离本发明精神和范围的前提下，本发明还会有各种变化和改进，这些变化和改进都落入要求保护的本发明范围内。本发明要求保护范围由所附的权利要求书及其等效物界定。

Claims (8)

  1. 一种基于智能移动终端的跌倒检测方法，其特征在于，包括以下步骤：
    步骤一：采集在身高、体重、年龄方面有良好代表性的人体活动数据，并构建数据集，然后对数据集进行预处理，并将预处理后的数据集中80％的人体活动数据作为训练集，将20％的人体活动数据作为测试集；
    步骤二：运用特征工程技术对数据集进行特征提取与分析，使用PCA降维技术对特征向量进行分析，选取优质特征进行下一步训练；
    步骤三：基于LSTM-FCN模型，然后设计一种改进的FallNet模型，在FCN网络和LSTM网络之前加上Batch Normalization层，Batch Normalization层对输入数据进行归一化，归一化的数据输入全卷机模块和LSTM模块，并在输入层加上全局最大池化和全局平均池化层，用于提取输入序列的幅度特征，最后在相应的每一个卷积激活模块后面都进行相同的操作，然后利用训练集对FallNet模型进行训练，并利用测试集对FallNet模型进行测试；
    步骤四：设计APP，APP采取基于短时间-长时间连续监测的方式，将训练好的FallNet模型内置在移动设备中，然后对采集的人体活动数据进行滑动窗处理，然后根据FallNet模型对采集的人体活动数据进行跌倒检测，并在移动设备中设置本地报警模块和远程报警模块，利用本地报警模块和远程报警模块对跌倒检测结果进行报警求助。
  2. 根据权利要求1所述的一种基于智能移动终端的跌倒检测方法，其特征在于：所述步骤一中对数据集进行预处理时，利用数据划分手段将数据集随机打乱。
  3. 根据权利要求1所述的一种基于智能移动终端的跌倒检测方法，其特征在于：所述LSTM模块为包含8个LSTM单元的循环神经网络层，LSTM模块根据输入的时间序列提取特征；针对输入的加速度数据，进行提取8个特征值，并与FCN网络的各级特征进行整合，用于输出层使用。
  4. 根据权利要求3所述的一种基于智能移动终端的跌倒检测方法，其特征在于：所述LSTM模块后面连接有参数丢失层，FallNet模型的每一轮训练中，参数丢失层随机丢失部分特征，实现重新训练网络、防止过拟 合。
  5. 根据权利要求1所述的一种基于智能移动终端的跌倒检测方法，其特征在于：所述步骤四中APP采取基于短时间-长时间连续监测的方式时，首先在短时间进行实时动作检测，然后进行长时间段的人体状态监测。
  6. 根据权利要求5所述的一种基于智能移动终端的跌倒检测方法，其特征在于：所述步骤四中进行长时间段的人体状态监测时，以动作监测序列作为数据判断依据，当发生跌倒动作，本地报警模块自动启动预报警程序，当在指定时间检测到用户继续正常活动，则取消报警，否则起动自动报警程序。
  7. 根据权利要求6所述的一种基于智能移动终端的跌倒检测方法，其特征在于：所述本地报警模块自动启动预报警程序的过程为：当检测到用户发生跌倒动作时，移动设备发出报警语音，进行本地求助；同时在预设置的等待时间内发出语音提示，询问用户跌倒是否属实，是否需要取消报警，如果用户选择取消，则取消报警，否则在指定时间结束时，立即报警；当检测到用户无动作时，同样进行语音询问，当用户按下取消键，则取消报警，否则认为用户已经跌倒并且属于危险状况，需要立即联系监护人和拨打急救电话。
  8. 根据权利要求7所述的一种基于智能移动终端的跌倒检测方法，其特征在于：所述远程报警模块用于当确认用户跌倒时，在本地报警模块报警求助的同时，远程报警模块同步将当前跌倒信息与位置发送到监护人手机并且拨打急救电话。
PCT/CN2020/137260 2020-04-20 2020-12-17 一种基于智能移动终端的跌倒检测方法 WO2021212883A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010309877.4 2020-04-20
CN202010309877.4A CN111436944B (zh) 2020-04-20 2020-04-20 一种基于智能移动终端的跌倒检测方法

Publications (1)

Publication Number Publication Date
WO2021212883A1 true WO2021212883A1 (zh) 2021-10-28

Family

ID=71654330

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/137260 WO2021212883A1 (zh) 2020-04-20 2020-12-17 一种基于智能移动终端的跌倒检测方法

Country Status (2)

Country Link
CN (1) CN111436944B (zh)
WO (1) WO2021212883A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114533046A (zh) * 2022-02-23 2022-05-27 成都华乾科技有限公司 一种基于csi信号的居家人员活动状态监测方法及***
CN114595748A (zh) * 2022-02-21 2022-06-07 南昌大学 一种用于跌倒防护***的数据分割方法
CN114842394A (zh) * 2022-05-17 2022-08-02 西安邮电大学 基于Swin Transformer的手术视频流程自动识别方法
CN116229581A (zh) * 2023-03-23 2023-06-06 珠海市安克电子技术有限公司 一种基于大数据的智能互联急救***
CN118015785A (zh) * 2024-04-07 2024-05-10 吉林大学 远程监测护理***及其方法

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112016619A (zh) * 2020-08-28 2020-12-01 西安科技大学 一种基于鞋垫的跌倒检测方法
CN112307287B (zh) * 2020-11-11 2022-08-02 国网山东省电力公司威海供电公司 基于云边协同架构的电力物联网数据分类处理方法及装置
CN112382052A (zh) * 2020-11-16 2021-02-19 南通市第一人民医院 一种基于互联网的患者跌倒报警方法及***
CN114067436B (zh) * 2021-11-17 2024-03-05 山东大学 一种基于可穿戴式传感器及视频监控的跌倒检测方法及***

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104125337A (zh) * 2014-07-22 2014-10-29 厦门美图移动科技有限公司 一种智能手机的跌倒检测和报警方法
CN108549841A (zh) * 2018-03-21 2018-09-18 南京邮电大学 一种基于深度学习的老人跌倒行为的识别方法
CN109394229A (zh) * 2018-11-22 2019-03-01 九牧厨卫股份有限公司 一种跌倒检测方法、装置及***
CN109670548A (zh) * 2018-12-20 2019-04-23 电子科技大学 基于改进lstm-cnn的多尺寸输入har算法
CN109726682A (zh) * 2018-12-29 2019-05-07 南京信息工程大学 一种面向弱标签传感器数据的人体动作识别方法
CN109979161A (zh) * 2019-03-08 2019-07-05 河海大学常州校区 一种基于卷积循环神经网络的人体跌倒检测方法
US20190287376A1 (en) * 2018-03-14 2019-09-19 Safely You Inc. System and Method for Detecting, Recording and Communicating Events in the Care and Treatment of Cognitively Impaired Persons
CN110659677A (zh) * 2019-09-10 2020-01-07 电子科技大学 一种基于可移动传感器组合设备的人体跌倒检测方法
CN112016619A (zh) * 2020-08-28 2020-12-01 西安科技大学 一种基于鞋垫的跌倒检测方法

Family Cites Families (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7733224B2 (en) * 2006-06-30 2010-06-08 Bao Tran Mesh network personal emergency response appliance
CN103308069B (zh) * 2013-06-04 2015-06-24 电子科技大学 一种跌倒检测装置及方法
US10485452B2 (en) * 2015-02-25 2019-11-26 Leonardo Y. Orellano Fall detection systems and methods
US10226204B2 (en) * 2016-06-17 2019-03-12 Philips North America Llc Method for detecting and responding to falls by residents within a facility
US20170173262A1 (en) * 2017-03-01 2017-06-22 François Paul VELTZ Medical systems, devices and methods
CA3073289A1 (en) * 2017-08-22 2019-02-28 Kinetyx Sciences Inc. Method and system for activity classification
CN108021888B (zh) * 2017-12-05 2021-09-24 电子科技大学 一种跌倒检测方法
CN110246300A (zh) * 2018-03-07 2019-09-17 深圳市智听科技有限公司 助听器的数据处理方法、装置
CN108683724A (zh) * 2018-05-11 2018-10-19 江苏舜天全圣特科技有限公司 一种智能儿童安全及步态健康监护***
CN109820515A (zh) * 2019-03-01 2019-05-31 中南大学 TensorFlow平台上基于LSTM神经网络的多传感跌倒检测的方法
CN110298278B (zh) * 2019-06-19 2021-06-04 中国计量大学 一种基于人工智能的地下停车库行人车辆监测方法
CN110321870B (zh) * 2019-07-11 2023-01-03 西北民族大学 一种基于lstm的掌静脉识别方法
CN110633736A (zh) * 2019-08-27 2019-12-31 电子科技大学 一种基于多源异构数据融合的人体跌倒检测方法
CN110420016B (zh) * 2019-08-28 2023-10-24 成都理工大学工程技术学院 一种运动员疲劳度的预测方法及***
CN110532966A (zh) * 2019-08-30 2019-12-03 深兰科技（上海）有限公司 一种基于分类模型进行跌倒识别的方法及设备
CN110738821A (zh) * 2019-09-27 2020-01-31 深圳市大拿科技有限公司 一种通过远程摄像告警方法及***

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104125337A (zh) * 2014-07-22 2014-10-29 厦门美图移动科技有限公司 一种智能手机的跌倒检测和报警方法
US20190287376A1 (en) * 2018-03-14 2019-09-19 Safely You Inc. System and Method for Detecting, Recording and Communicating Events in the Care and Treatment of Cognitively Impaired Persons
CN108549841A (zh) * 2018-03-21 2018-09-18 南京邮电大学 一种基于深度学习的老人跌倒行为的识别方法
CN109394229A (zh) * 2018-11-22 2019-03-01 九牧厨卫股份有限公司 一种跌倒检测方法、装置及***
CN109670548A (zh) * 2018-12-20 2019-04-23 电子科技大学 基于改进lstm-cnn的多尺寸输入har算法
CN109726682A (zh) * 2018-12-29 2019-05-07 南京信息工程大学 一种面向弱标签传感器数据的人体动作识别方法
CN109979161A (zh) * 2019-03-08 2019-07-05 河海大学常州校区 一种基于卷积循环神经网络的人体跌倒检测方法
CN110659677A (zh) * 2019-09-10 2020-01-07 电子科技大学 一种基于可移动传感器组合设备的人体跌倒检测方法
CN112016619A (zh) * 2020-08-28 2020-12-01 西安科技大学 一种基于鞋垫的跌倒检测方法

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114595748A (zh) * 2022-02-21 2022-06-07 南昌大学 一种用于跌倒防护***的数据分割方法
CN114595748B (zh) * 2022-02-21 2024-02-13 南昌大学 一种用于跌倒防护***的数据分割方法
CN114533046A (zh) * 2022-02-23 2022-05-27 成都华乾科技有限公司 一种基于csi信号的居家人员活动状态监测方法及***
CN114533046B (zh) * 2022-02-23 2024-05-07 成都华乾科技有限公司 一种基于csi信号的居家人员活动状态监测方法及***
CN114842394A (zh) * 2022-05-17 2022-08-02 西安邮电大学 基于Swin Transformer的手术视频流程自动识别方法
CN114842394B (zh) * 2022-05-17 2024-04-16 西安邮电大学 基于Swin Transformer的手术视频流程自动识别方法
CN116229581A (zh) * 2023-03-23 2023-06-06 珠海市安克电子技术有限公司 一种基于大数据的智能互联急救***
CN116229581B (zh) * 2023-03-23 2023-09-19 珠海市安克电子技术有限公司 一种基于大数据的智能互联急救***
CN118015785A (zh) * 2024-04-07 2024-05-10 吉林大学 远程监测护理***及其方法
CN118015785B (zh) * 2024-04-07 2024-06-07 吉林大学 远程监测护理***及其方法

Also Published As

Publication number Publication date
CN111436944A (zh) 2020-07-24
CN111436944B (zh) 2021-09-28

Similar Documents

Publication Publication Date Title
WO2021212883A1 (zh) 一种基于智能移动终端的跌倒检测方法
CN104125337B (zh) 一种智能手机的跌倒检测和报警方法
US20170372582A1 (en) Device for monitoring and preventing slump and processing method thereof
Shany et al. Are we stumbling in our quest to find the best predictor? Over‐optimism in sensor‐based models for predicting falls in older adults
CN104771177B (zh) 一种基于机器学习的跌倒检测***及方法
CN110148405B (zh) 语音指令处理方法及装置、电子设备及存储介质
US20210064141A1 (en) System for detecting a signal body gesture and method for training the system
WO2015009951A1 (en) Fall detection using machine learning
CN107609477A (zh) 一种基于深度学习与智能手环相结合的摔倒检测方法
US20230190188A1 (en) Method and system for remote transdermal alcohol monitoring
De Falco et al. Evaluation of artificial intelligence techniques for the classification of different activities of daily living and falls
JP2020527063A (ja) センサの動作ステータスを検出するための方法および状態機械システム
CN107105092A (zh) 一种基于动态时间规整的人体跌倒识别方法
CN112016619A (zh) 一种基于鞋垫的跌倒检测方法
CN113516828A (zh) 溺水监测方法、装置、设备与计算机可读存储介质
CN117807551B (zh) 基于智能戒指的心率异常捕获方法及***
Dham et al. Mental stress detection using artificial intelligence models
CN118035904A (zh) 一种基于多频段数据融合的振摆监测故障诊断方法及***
WO2022006659A1 (en) Method and apparatus for detecting fall events
CN116763299A (zh) 基于可穿戴传感器的跌倒检测***及方法
CN113096792A (zh) 一种基于机器学习的智能健康监测预警方法及***
CN108648417A (zh) 基于深度学习的树莓派老人摔倒检测***
CN113671489B (zh) 状态提醒方法及装置、电子设备和计算机可读存储介质
CN112386249B (zh) 跌倒检测方法及装置、设备、存储介质
CN113095153A (zh) 基于深度残差网络的移动终端人类情境识别方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20931862

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20931862

Country of ref document: EP

Kind code of ref document: A1