WO2021115064A1 - 基于可穿戴传感器的健身运动识别方法 - Google Patents

基于可穿戴传感器的健身运动识别方法 Download PDF

Info

Publication number
WO2021115064A1
WO2021115064A1 PCT/CN2020/129525 CN2020129525W WO2021115064A1 WO 2021115064 A1 WO2021115064 A1 WO 2021115064A1 CN 2020129525 W CN2020129525 W CN 2020129525W WO 2021115064 A1 WO2021115064 A1 WO 2021115064A1
Authority
WO
WIPO (PCT)
Prior art keywords
fitness exercise
method based
recognition method
signal
fitness
Prior art date
Application number
PCT/CN2020/129525
Other languages
English (en)
French (fr)
Inventor
孙方敏
田佳承
李烨
Original Assignee
中国科学院深圳先进技术研究院
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 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Publication of WO2021115064A1 publication Critical patent/WO2021115064A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
    • G01C23/00—Combined instruments indicating more than one navigational value, e.g. for aircraft; Combined measuring devices for measuring two or more variables of movement, e.g. distance, speed or acceleration
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Definitions

  • the invention belongs to the field of computer technology and application technology, and relates to a fitness exercise recognition method based on a wearable sensor.
  • the human body behavior recognition method based on video equipment realizes human body behavior recognition by acquiring the image information of human motion and analyzing its image sequence.
  • video acquisition equipment is required to obtain target video information.
  • Such equipment is usually large, fixed in location, high power consumption, large calculations, and poor anti-interference ability, which may cause unforeseen environments. Factors interfere with the data, which makes it only suitable for fixed occasions and cannot adapt to long-term continuous human behavior records.
  • the sensor is integrated into the sports equipment.
  • some manufacturers have introduced similar equipment.
  • users can scan the QR code to activate the sensor to record the data.
  • the movement information is sent to the user's mobile phone.
  • HAR Human Activity Recognition
  • Environment-based sensors can generally be used in fixed-scene environments, such as homes and gyms. Generally, sensors are placed on some objects, and when the user uses these sensors, the sensors are activated to record user behavior information. Although the environmental sensor-based method has been used in some scenarios, such a system layout still requires a large number of sensors. The cost of the equipment is higher than that of ordinary equipment, and because of the need for power or built-in power, there are certain safety risks, and The outdoor behavior of users cannot be monitored, and it is more difficult to distinguish specific users.
  • Mobile wearable sensing devices can monitor the user's indoor and outdoor behaviors at any time through sensor nodes worn on the human body. Secondly, because the sensor equipment is independently owned by itself, there is no need to worry about privacy leakage, and wearable sensors can integrate many kinds of components to collect different signals, through which different body movements of the human body and the health of the body can be analyzed. The status is detected simultaneously.
  • a Chinese patent (name: a method for user behavior recognition based on smart mobile device sensors, application number: CN201910347816.4) discloses a method for user behavior recognition based on smart mobile device sensors. The method obtains acceleration and angular velocity data. , It is imaged and processed according to the way of image processing, but the recognition action is relatively simple, which is not enough to support the current real needs.
  • the Chinese patent (name: human behavior recognition method based on convolutional neural network and recurrent neural network, application number: CN201910580116.X) discloses a human behavior recognition method based on convolutional neural network and recurrent neural network.
  • the data In addition to the sensor data, the data also collects the RGB video of the scene. It is easy to be affected by light, obstacles, etc. during the collection, and an ideal environment is required to collect data.
  • the purpose of the present invention is to provide a wearable sensor-based fitness exercise recognition method, which uses the acceleration and angular velocity signals collected by the motion sensor integrated in the wearable device during the fitness exercise of the individual to extract exercise information, So as to realize the recognition of fitness exercises.
  • Step 1 Obtain the inertial sensor signal during the fitness exercise and perform preprocessing
  • Step 2 Perform window segmentation on the signal preprocessed in Step 1;
  • Step 3 Perform feature extraction on the signal segmented in Step 2;
  • Step 4 normalize the feature data extracted in step 3;
  • Step 5 Perform feature dimensionality reduction on the features processed in step 4;
  • Step 6 Identify the features processed in step 5.
  • the acquisition process of fitness exercise signals in step 1 is: use the accelerometer and gyroscope built in the wearable device to acquire the acceleration signals and angular velocity signals of the fitness exercise;
  • the process of signal preprocessing is: linear interpolation and desiccation processing on the collected acceleration signal and angular velocity signal.
  • step 2 the window segmentation method includes segmentation based on sliding windows, window segmentation based on event definitions, and window segmentation based on action definitions.
  • each window has a 50% overlap of information.
  • step 3 includes maximum value, minimum value, mean value, variance, skewness, kurtosis, maximum peak value and energy of the discrete Fourier transform spectrogram;
  • ⁇ represents the mean
  • ⁇ 2 represents the variance
  • S max represents the maximum value in the vector
  • S min represents the minimum value in the vector
  • ske represents the skewness
  • kur represents the kurtosis
  • S DFT (k) represents the discrete Fourier transform
  • E represents energy
  • step 4 The specific process of step 4 is as follows: the extracted features are formed into a feature vector, and the vector is normalized to the interval [0,1] for the training of the classifier;
  • step 5 principal component analysis is used for dimensionality reduction.
  • step 6 a hierarchical method based on neural network is used to identify the behavior of people exercising in the gym.
  • the beneficial effect of the present invention is that the fitness exercise recognition method based on the wearable sensor of the present invention uses the accelerometer and gyroscope integrated in the wearable sensor to separately collect the acceleration signal and angular velocity signal of the human body during fitness exercise. Through preprocessing, window segmentation and feature extraction of acceleration and angular velocity signals, dimensionality reduction processing and classification modeling of the extracted features are implemented to realize the recognition of fitness exercises.
  • Fig. 1 is a flowchart of a fitness exercise recognition method based on a wearable sensor of the present invention.
  • the fitness exercise recognition method based on the wearable sensor of the present invention specifically includes the following steps:
  • Step 1 Obtain the inertial sensor signal during the fitness exercise and perform preprocessing
  • Linear interpolation Because the built-in acceleration sensor in the wearable device has worse performance than the independent acceleration sensor, the sampling clock is unstable, resulting in unequal time intervals between consecutive acceleration sampling points. To solve this problem, linear interpolation is used The method to ensure that the time interval between two sample points is fixed.
  • the motion signal captured by the accelerometer and gyroscope contains a lot of noise, which may be caused by the unfixed position of the sensor, the resetting of exercises during fitness, and the shaking of the body when overcoming weights.
  • the moving average filter is a low-pass filter, which can effectively reduce the influence of random interference. Therefore, the present invention proposes a fifth-order moving average filter algorithm to eliminate noise and reduce signal noise caused by the collection environment.
  • Step 2 Perform window segmentation on the signal preprocessed in step 1; perform sliding window segmentation technology on the preprocessed data stream, set the window length to 2 seconds, and each window has a 50% overlap of information.
  • segmentation techniques mainly include three methods: sliding window-based segmentation, event-based window segmentation, and action-defined window segmentation.
  • Event-based window segmentation is to divide the data stream according to different events.
  • the start and end of each window represent the beginning and end of an event. This method needs to use related algorithms to determine the start and end of the event, so it is rarely used in action recognition.
  • Window segmentation based on action definition divides the data stream into windows of different time lengths according to different action types, and each window represents an action. This method is mainly based on the difference between different action signals to segment, and it is difficult to apply to real-time action recognition systems.
  • Sliding window segmentation technology refers to the use of fixed-length windows to segment the data stream.
  • the length of the window is set to 1 second, 3 seconds, 6 seconds, 12 seconds, etc.
  • the data of adjacent windows can be partially overlapped or completely disjoint.
  • the purpose of data overlap is to be more accurate Process the transition between actions.
  • the present invention uses sliding window segmentation technology for processing, the window length is set to 2 seconds, and there is 50% signal overlap between adjacent windows, which can effectively avoid the loss of information.
  • Step 3 Perform feature extraction on the signal segmented in Step 2;
  • the feature extraction is divided into two parts: the bottom layer (ankle, thigh) and the top layer (waist, wrist, arm).
  • the bottom layer ankle, thigh
  • the top layer waist, wrist, arm
  • the average value, variance and energy of the bottom layer are extracted as features.
  • the movements of the upper body are more complex and similar, so more features need to be extracted, and finally the maximum value, minimum value, mean value, variance, skewness, kurtosis, and the 5 maximum peaks of the discrete Fourier transform spectrogram are selected. Energy, etc. as a feature.
  • the extraction of data features in the present invention includes two parts: time domain and frequency domain features.
  • the time domain features include maximum, minimum, mean, variance, skewness, and kurtosis; the frequency domain features mainly select the five maximum peaks and energy of the discrete Fourier transform spectrum.
  • ⁇ represents the mean
  • ⁇ 2 represents the variance
  • S max represents the maximum value in the vector
  • S min represents the minimum value in the vector
  • ske represents the skewness
  • kur represents the kurtosis
  • S DFT (k) represents the discrete Fourier transform ( Discrete Fourier Transform (DFT) is the peak value of the k-th element
  • E represents energy.
  • Step 4 normalize the feature data extracted in step 3;
  • the extracted features are formed into a feature vector, and the vector is normalized to the interval [0,1] through a formula for the training of the classifier.
  • the present invention uses normalization to process the data so that it is limited to the range of [0,1].
  • the formula is as follows:
  • Step 5 Perform feature dimensionality reduction on the features processed in step 4;
  • the dimensionality of the composed feature vector is high, so it is necessary to reduce the dimensionality of the features
  • the present invention uses the principal component analysis method (PCA) to reduce the dimensionality of the feature vector obtained as described above.
  • PCA principal component analysis method
  • Step 6 Identify the features processed in step 5.
  • the back propagation (BP) neural network is used to train and classify the samples.
  • the features mean, variance, energy
  • the features are extracted from the ankle and thigh node data
  • only train one BP neural network to classify the four lower body states, so that different concurrent actions can be divided into four groups, effectively reducing the decision boundary Complexity.
  • On the top layer of the system extract the features (12 features) of the data of the wrist, arm and waist nodes and perform dimensionality reduction, and design the top-level neural network corresponding to different lower limb states to identify upper body movements and infer the final fitness movements .
  • the characteristics of the fitness exercise recognition method based on the wearable sensor of the present invention are that the accelerometer and gyroscope integrated in the wearable sensor are used to separately collect the acceleration signal and angular velocity signal of the human body during fitness exercise.
  • window segmentation and feature extraction of acceleration and angular velocity signals, dimensionality reduction processing and classification modeling of the extracted features are implemented to realize the recognition of fitness exercises.
  • the main contents include: an effective fixed-length sliding window segmentation method is proposed to divide the sensor data stream.
  • the window length is 2 seconds, and there is 50% signal overlap in adjacent windows, which can effectively avoid the loss of information;
  • An effective neural network-based hierarchical recognition method realizes the recognition of concurrent upper and lower body actions during fitness;
  • An effective feature extraction method is proposed, which selects different time domains of exercise cycles according to different layers And frequency domain characteristics, can better reflect the characteristics of the runtime; finally use several common classification techniques including least squares, naive Bayes and k-nearest neighbor algorithm test comparison, and summarize the most effective for fitness exercise recognition method.

Abstract

一种基于可穿戴传感器的健身运动识别方法，具体包括如下步骤：步骤1，获取健身运动过程中的惯性传感信号并进行预处理；步骤2，对步骤1预处理后的信号进行窗口分割；步骤3，对步骤2分割后的信号进行特征提取；步骤4，对步骤3提取的特征数据进行归一化处理；步骤5，对步骤4处理后的特征进行特征降维；步骤6，对经步骤5处理后的特征进行识别。该方法利用个人在健身运动的过程中，集成在可穿戴设备中的运动传感器采集到的加速度和角速度信号，提取运动信息，从而实现对健身运动的识别。

Description

基于可穿戴传感器的健身运动识别方法 技术领域
本发明属于计算机技术与应用技术领域，涉及一种基于可穿戴传感器的健身运动识别方法。
背景技术
传统的人体行为识别技术主要有两种，一是通过录像、拍照的方式获取人体行为的视频、图像数据，二是通过基于环境传感器的方式来获取数据，三是通过可穿戴设备来获取数据。基于视频设备的人体行为识别方法通过获取人体运动的的图像信息，对其图像序列进行分析从而实现人体行为识别。但在实际的运用中，获取目标视频信息需要视频采集设备，而这类设备通常体积较大、位置固定、功耗较高、计算量大，同时抗干扰能力差，可能会存在无法预料的环境因素对数据造成干扰，导致其只适用了固定的场合，无法适应长时持续性的人体行为记录。
基于环境传感器的方式来获取数据，则是将传感器集成到运动设备上，目前有的厂家推出相类似的设备，使用者在使用设备时，通过二维码扫码，激活传感器记录数据，并将运动的信息发送到使用者手机里。
近年来，随着微电子机械制造行业技术的进步，各种类型的传感器，诸如加速度计、陀螺仪、磁力计等，能够采集个人运动的信息，同时具备良好的便携性，具有低功耗、抗环境干扰等功能，成为了实现长时间连续性记录人体行为信息的首选。
人体行为识别(Human Activity Recognition，HAR)的研究可以分为三种主要的方法：基于机器视觉、基于环境传感器的行为识别、基于可穿戴传感器的行为识别。
基于机器视觉：在人体行为识别的早期研究中，大多使用的都是基于机器视觉的方法。通常这些系统采用一个或多个摄像设备对人体进行检测，在获得视频信息后，通过这些信息序列，提取人的身体信息，并采用机器学习或模型推理识别出人体模型的动作。然而，这一方法在实际使用中容易收到环境因素的干扰，并且场景固定、设备昂贵、安全性不足这一系列问题。
基于环境传感器：基于环境传感器一般可以应用于固定场景的环境中，如家庭、健身房这些场景。通常，在一些物体上放置传感器，当用户使用到这些传感器的时候，传感器被激活，从而实现记录用户行为行为信息。尽管基于环境传感器的方法在一些场景得到了使用，但是这样的系统布置仍需要大量的传感器，设备的费用相对于普通的设备费用高昂，并且因 为需要通电或内置电源，有一定的安全隐患，并且无法监测到使用者的户外行为，并且比较难区分出具体的用户。
基于可穿戴传感器的行为识别：移动可穿戴传感设备通过佩戴在人体躯体上的传感器节点，可以在任意时间监测用户在室内、外的行为活动。其次，因为传感器设备是自己独立拥有的，所以不用担心隐私泄漏的问题，并且可穿戴传感器可以集成许多种元件用来收集不同的信号，通过这些信号可以分析人体不同的肢体动作和对身体的健康状况同步进行检测。
中国专利(名称：基于多源数据融合的健身运动识别方法及系统，申请号:CN201710525603.7)公开了一种基于多源数据融合的健身运动识别方法及系统，它在数据源的获取上需要融合基于环境传感器(如健身器械)上的数据，虽然在一定程度上能提高识别的准确性，但越多的动作需要的能收集数据的器械也越多，费用昂贵，适用性不够好。
中国专利(名称：一种基于智能移动设备传感器的用户行为识别方法，申请号：CN201910347816.4)公开了一种基于智能移动设备传感器的用户行为识别方法，该方法在获得了加速度和角速度数据后，将其图像化后按照图像处理的方式进行处理，但是其识别的动作较为简单，不足以支撑当下的现实需求。
中国专利(名称：基于卷积神经网络和循环神经网络的人体行为识别方法，申请号：CN201910580116.X)公开了一种基于卷积神经网络和循环神经网络的人体行为识别方法，该方法收集的数据除了传感器的数据外，还收集了场景的RGB视频，在采集时很容易收到光照，障碍物等影响，需要较为理想化的环境收集数据。
发明内容
本发明的目的是提供一种基于可穿戴传感器的健身运动识别方法，该方法利用个人在健身运动的过程中，集成在可穿戴设备中的运动传感器采集到的加速度和角速度信号，提取运动信息，从而实现对健身运动的识别。
本发明所采用的技术方案是，基于可穿戴传感器的健身运动识别方法，具体包括如下步骤：
步骤1，获取健身运动过程中的惯性传感信号并进行预处理；
步骤2，对步骤1预处理后的信号进行窗口分割；
步骤3，对步骤2分割后的信号进行特征提取；
步骤4，对步骤3提取的特征数据进行归一化处理；
步骤5，对步骤4处理后特征进行特征降维；
步骤6，对经步骤5处理后的特征进行识别。
本发明的特征还在于：
步骤1中健身运动信号的获取过程为：利用可穿戴设备中内置的加速度计和陀螺仪来获 取健身运动的加速度信号和角速度信号；
信号预处理的过程为：对采集到的加速度信号和角速度信号进行线性插值和去燥处理。
步骤2的具体过程为：所述窗口分割方法包括基于滑动窗口的分割、基于事件定义的窗口分割和基于动作定义的窗口分割。
步骤2中每个窗口有50％信息的重叠。
步骤3中的特征包括最大值、最小值、均值、方差、偏度、峰度、离散傅里叶变换频谱图的最大峰值和能量；
各特征的具体计算公式如下：
Figure PCTCN2020129525-appb-000001
Figure PCTCN2020129525-appb-000002
S max＝max(s)        (3)；
S min＝min(s)        (4)；
Figure PCTCN2020129525-appb-000003
Figure PCTCN2020129525-appb-000004
Figure PCTCN2020129525-appb-000005
Figure PCTCN2020129525-appb-000006
其中μ表示均值、σ 2表示方差、S max表示向量中的最大值、S min表示向量中的最小值、ske表示偏度、kur表示峰度、S DFT(k)表示离散傅里叶变换的第k个元素的峰值、E表示能量。
步骤4的具体过程如下：将提取的特征组成特征向量，将向量归一化到[0,1]区间，用于分类器的训练；
根据如下公式(9)对数据进行归一化处理：
Figure PCTCN2020129525-appb-000007
步骤5中采用主成分分析法进行降维。
步骤6中采用基于神经网络的分层方法对人在健身房运动的行为进行识别。
本发明的有益效果是：本发明基于可穿戴传感器的健身运动识别方法，利用可穿戴传感器内集成的加速度计和陀螺仪分别采集人体在健身运动时的加速度信号、角速度信号。通过对加速度、角速度信号进行预处理、窗口分割、特征提取，对提取的特征进行降维处理、分类建模等实现对健身运动的识别。
附图说明
图1是本发明基于可穿戴传感器的健身运动识别方法的流程图。
具体实施方式
下面结合附图和具体实施方式对本发明进行详细说明。
本发明基于可穿戴传感器的健身运动识别方法，如图1所示，具体包括如下步骤：
步骤1，获取健身运动过程中的惯性传感信号并进行预处理；
利用可穿戴设备中内置的加速度计和陀螺仪来获取健身运动的加速度(Acc)信号和角速度(Gyro)信号，对采集到的数据信号分别进行线性插值和去噪。
线性插值：由于可穿戴设备中内置的加速度传感器比独立加速度传感器性能较差，采样时钟不稳定，导致相邻的连续加速度采样点之间的时间间隔不相等，为解决这一问题，采用线性插值的方法来确保两样本点之间的时间间隔是固定的。
去除噪音：加速度计和陀螺仪捕获的运动信号含有大量的噪声，这可能是由于传感器位置不固定、健身时动作复位和克服重量做功时身体的抖动造成的。移动平均滤波器是一个低通滤波器，可以有效地减少随机干扰的影响，因此本发明提出了5阶移动平均滤波算法来消除噪声，降低由采集环境带来的信号噪声。
步骤2，对步骤1预处理后的信号进行窗口分割；对预处理后的数据流进行滑动窗口分割技术进行处理，窗口长度设为2秒，并且每个窗口有50％信息的重叠。
对于传感器数据流，在动作识别前需要将其划分为一个个小的时间窗口。常用的分割技术主要包括基于滑动窗口的分割、基于事件定义的窗口分割和基于动作定义的窗口分割等三种方法。
基于事件定义的窗口分割是按照不同的事件对数据流进行划分，每个窗口的起点和终点分别代表一个事件的开始和结束。该方法需要借助相关算法确定事件的起点和终点，因此，在动作识别中很少被使用。
基于动作定义的窗口分割是根据不同的动作类型，将数据流划分为不同时间长度的窗口，每个窗口代表一个动作。该方法主要是根据不同动作信号间的差异来进行分割，难以应用到实时动作识别系统中。
滑动窗口分割技术是指采用固定长度的窗口对数据流进行分割。在不同的人体动作识别研究中，窗口的长度被设置为1秒、3秒、6秒、12秒等，相邻窗口的数据可以部分重叠或完全不相交，数据重叠的目的是为了更准确地处理动作之间的转换。
本发明采用滑动窗口分割技术进行处理，窗口长度设置为2秒，相邻的窗口间存在50％信号的交叠，可以有效避免信息的丢失。
步骤3，对步骤2分割后的信号进行特征提取；
特征的提取分为两个部分：底层(脚踝、大腿)、顶层(腰、手腕、手臂)。健身运动中下身的动作较少，大致可为分别坐姿、站立、运动三个状态，因此提取底层的均值、方差和能量作为特征。相对地，上半身的动作较为复杂、相似，因此需提取较多的特征，最终选择最大值、最小值、均值、方差、偏度、峰度、离散傅里叶变换频谱图的5个最大峰值和能量等作为特征。
对于每一个时间窗口，它所包含的数据都可以用一个N*1维的向量S＝[S 1,S 2,…,S N] T来表示。本发明对数据特征的提取包括时域和频域特征两部分。时域特征有最大值、最小值、均值、方差、偏度和峰度；频域特征选取的主要是离散傅里叶变换谱图的5个最大峰值和能量。
各特征的计算公式如下：
Figure PCTCN2020129525-appb-000008
Figure PCTCN2020129525-appb-000009
S max＝max(s)         (3)；
S min＝min(s)          (4)；
Figure PCTCN2020129525-appb-000010
Figure PCTCN2020129525-appb-000011
Figure PCTCN2020129525-appb-000012
Figure PCTCN2020129525-appb-000013
其中μ表示均值、σ 2表示方差、S max表示向量中的最大值、S min表示向量中的最小值、ske表示偏度、kur表示峰度、S DFT(k)表示离散傅里叶变换(Discrete Fourier Transform,DFT)的第k个元素的峰值，E表示能量。
步骤4，对步骤3提取的特征数据进行归一化处理；
将提取的特征组成特征向量，通过公式将向量归一化到[0,1]区间，用于分类器的训练。
不同的评价指标(即特征向量中的不同特征就是所述的不同评价指标)往往具有不同的量纲和量纲单位，这样的情况会影响到数据分析的结果，为了消除指标之间的量纲影响，需要进行数据标准化处理，以解决数据指标之间的可比性。原始数据经过数据标准化处理后，各指标处于同一数量级，适合进行综合对比评价。
本发明为了消除奇异样本数据导致的不良影响，为了提升模型的收敛速度和模型的精度，使用归一化处理数据，让其被限定在[0,1]范围内。公式如下：
Figure PCTCN2020129525-appb-000014
步骤5，对步骤4处理后特征进行特征降维；
在顶层因为选取的特征较多，组成的特征向量维度高，因此需要对特征进行降维；
在基于可穿戴传感器数据的人体动作识别中，当提取的特征数量过多时，直接对高维特征进行分类会遇到许多问题，包括会产生大量的计算、数据不够直观以及数据可视性较差等。而在特征集合中，这些特征之间有些是相互关联的，并且在一定程度上存在着信息的重叠。为防止计算量过大，本发明将上述述取得的特征向量使用主成分分析方法(PCA)降维。
步骤6，对经步骤5处理后的特征进行识别。
在获得所提取的特征向量后，使用前馈反向(Back propagation，BP)神经网络对样本进行训练和分类。在系统的底层，提取脚踝和大腿节点数据的特征(均值、方差、能量)，只训练一个BP神经网络来分类四种下身状态，这样可以将不同的并发动作分为四组，有效降低决策边界的复杂度。在系统的顶层，提取手腕、手臂和腰部节点的数据的特征(12个特征)并进行降维，并对应不同的下肢状态设计顶层的神经网络，来识别上身动作，并推断出最终的健身动作。
本发明采用基于神经网络的分层方法对人在健身房运动的行为进行识别。在对数据进行了降维处理，归一化处理之后，为了验证分层处理的有效性，本发明也采用几种常用的分类技术，通过结合分层、单层的方法进行比较，包括最小二乘法、朴素贝叶斯和K近邻方法(k＝1)。
本发明基于可穿戴传感器的健身运动识别方法的特点为，利用可穿戴传感器内集成的加速度计和陀螺仪分别采集人体在健身运动时的加速度信号、角速度信号。通过对加速度、角速度信号进行预处理、窗口分割、特征提取，对提取的特征进行降维处理、分类建模等实现对健身运动的识别。主要内容包括：提出了一种有效的固定长度的滑动窗口分割方法对传感器数据流进行划分，窗口长度为2秒，且相邻窗口存在50％的信号叠加，可以有效避免信息的丢失；提出一种有效的基于神经网络的分层识别方法，实现了对人在健身中上身和下身并发动作的识别；提出了一种有效的特征提取方法，根据分层的不同选取不同的运动周期的时域和频域特征，可以更好地反映运行时的特性；最后利用几种常见的分类技术包括最小二乘法、朴素贝叶斯和k近邻算法测试比较，并总结出用于健身运动识别最有效的方法。

Claims (8)

  1. 基于可穿戴传感器的健身运动识别方法，其特征在于：具体包括如下步骤：
    步骤1，获取健身运动过程中的惯性传感信号并进行预处理；
    步骤2，对步骤1预处理后的信号进行窗口分割；
    步骤3，对步骤2分割后的信号进行特征提取；
    步骤4，对步骤3提取的特征数据进行归一化处理；
    步骤5，对步骤4处理后特征进行特征降维；
    步骤6，对经步骤5处理后的特征进行识别。
  2. 根据权利要求1所述的基于可穿戴传感器的健身运动识别方法，其特征在于：所述步骤1中健身运动信号的获取过程为：利用可穿戴设备中内置的加速度计和陀螺仪来获取健身运动的加速度信号和角速度信号；
    信号预处理的过程为：对采集到的加速度信号和角速度信号进行线性插值和去燥处理。
  3. 根据权利要求2所述的基于可穿戴传感器的健身运动识别方法，其特征在于：所述步骤2的具体过程为：所述窗口分割方法包括基于滑动窗口的分割、基于事件定义的窗口分割和基于动作定义的窗口分割。
  4. 根据权利要求2所述的基于可穿戴传感器的健身运动识别方法，其特征在于：所述
    步骤2中每个窗口有50％信息的重叠。
  5. 根据权利要求1所述的基于可穿戴传感器的健身运动识别方法，其特征在于：所述步骤3中的特征包括最大值、最小值、均值、方差、偏度、峰度、离散傅里叶变换频谱图的最大峰值和能量；
    各特征的具体计算公式如下：
    Figure PCTCN2020129525-appb-100001
    Figure PCTCN2020129525-appb-100002
    S max＝max(s)    (3)；
    S min＝min(s)    (4)；
    Figure PCTCN2020129525-appb-100003
    Figure PCTCN2020129525-appb-100004
    Figure PCTCN2020129525-appb-100005
    Figure PCTCN2020129525-appb-100006
    其中μ表示均值、σ 2表示方差、S max表示向量中的最大值、S min表示向量中的最小值、ske表示偏度、kur表示峰度、S DFT(k)表示离散傅里叶变换的第k个元素的峰值、E表示能量。
  6. 根据权利要求1所述的基于可穿戴传感器的健身运动识别方法，其特征在于：所述步骤4的具体过程如下：将提取的特征组成特征向量，将向量归一化到[0,1]区间，用于分类器的训练；
    根据如下公式(9)对数据进行归一化处理：
    Figure PCTCN2020129525-appb-100007
  7. 根据权利要求1所述的基于可穿戴传感器的健身运动识别方法，其特征在于：所述步骤5中采用主成分分析法进行降维。
  8. 根据权利要求1所述的基于可穿戴传感器的健身运动识别方法，其特征在于：所述步骤6中采用基于神经网络的分层方法对人在健身房运动的行为进行识别。
PCT/CN2020/129525 2019-12-10 2020-11-17 基于可穿戴传感器的健身运动识别方法 WO2021115064A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911257652.2 2019-12-10
CN201911257652.2A CN111089604B (zh) 2019-12-10 2019-12-10 基于可穿戴传感器的健身运动识别方法

Publications (1)

Publication Number Publication Date
WO2021115064A1 true WO2021115064A1 (zh) 2021-06-17

Family

ID=70394962

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/129525 WO2021115064A1 (zh) 2019-12-10 2020-11-17 基于可穿戴传感器的健身运动识别方法

Country Status (2)

Country Link
CN (1) CN111089604B (zh)
WO (1) WO2021115064A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117574133A (zh) * 2024-01-11 2024-02-20 湖南工商大学 一种不安全生产行为识别方法及相关设备

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111089604B (zh) * 2019-12-10 2021-09-07 中国科学院深圳先进技术研究院 基于可穿戴传感器的健身运动识别方法
CN112633467A (zh) * 2020-11-25 2021-04-09 超越科技股份有限公司 一种基于猫眼连接改进lstm模型的人体行为识别方法
CN113591552A (zh) * 2021-06-18 2021-11-02 新绎健康科技有限公司 一种基于步态加速度数据进行身份识别的方法及系统
CN117084671B (zh) * 2023-10-19 2024-04-02 首都医科大学宣武医院 一种基于陀螺仪信号的运动评估系统

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2010008900A1 (en) * 2008-06-24 2010-01-21 Dp Technologies, Inc. Program setting adjustments based on activity identification
CN109933202A (zh) * 2019-03-20 2019-06-25 深圳大学 一种基于骨传导的智能输入方法和系统
WO2019122168A1 (en) * 2017-12-21 2019-06-27 Yoti Holding Limited Biometric user authentication
CN110245718A (zh) * 2019-06-21 2019-09-17 南京信息工程大学 一种基于联合时域频域特征的人体行为识别方法
CN110334573A (zh) * 2019-04-09 2019-10-15 北京航空航天大学 一种基于密集连接卷积神经网络的人体运动状态判别方法
CN110532898A (zh) * 2019-08-09 2019-12-03 北京工业大学 一种基于智能手机多传感器融合的人体活动识别方法
CN111089604A (zh) * 2019-12-10 2020-05-01 中国科学院深圳先进技术研究院 基于可穿戴传感器的健身运动识别方法

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR102387378B1 (ko) * 2014-10-07 2022-04-15 삼성전자주식회사 보행 동작 인식을 위한 장치 및 방법
US10477354B2 (en) * 2015-02-20 2019-11-12 Mc10, Inc. Automated detection and configuration of wearable devices based on on-body status, location, and/or orientation
CN108549856B (zh) * 2018-04-02 2021-04-30 上海理工大学 一种人体动作和路况识别方法
CN108764282A (zh) * 2018-04-19 2018-11-06 中国科学院计算技术研究所 一种类别增量行为识别方法和系统
CN109086667A (zh) * 2018-07-02 2018-12-25 南京邮电大学 基于智能终端的相似活动识别方法

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2010008900A1 (en) * 2008-06-24 2010-01-21 Dp Technologies, Inc. Program setting adjustments based on activity identification
WO2019122168A1 (en) * 2017-12-21 2019-06-27 Yoti Holding Limited Biometric user authentication
CN109933202A (zh) * 2019-03-20 2019-06-25 深圳大学 一种基于骨传导的智能输入方法和系统
CN110334573A (zh) * 2019-04-09 2019-10-15 北京航空航天大学 一种基于密集连接卷积神经网络的人体运动状态判别方法
CN110245718A (zh) * 2019-06-21 2019-09-17 南京信息工程大学 一种基于联合时域频域特征的人体行为识别方法
CN110532898A (zh) * 2019-08-09 2019-12-03 北京工业大学 一种基于智能手机多传感器融合的人体活动识别方法
CN111089604A (zh) * 2019-12-10 2020-05-01 中国科学院深圳先进技术研究院 基于可穿戴传感器的健身运动识别方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117574133A (zh) * 2024-01-11 2024-02-20 湖南工商大学 一种不安全生产行为识别方法及相关设备
CN117574133B (zh) * 2024-01-11 2024-04-02 湖南工商大学 一种不安全生产行为识别方法及相关设备

Also Published As

Publication number Publication date
CN111089604A (zh) 2020-05-01
CN111089604B (zh) 2021-09-07

Similar Documents

Publication Publication Date Title
WO2021115064A1 (zh) 基于可穿戴传感器的健身运动识别方法
Wang et al. Fall detection based on dual-channel feature integration
Mubashir et al. A survey on fall detection: Principles and approaches
Gupta et al. A survey on human activity recognition and classification
Bennett et al. Inertial measurement unit-based wearable computers for assisted living applications: A signal processing perspective
Zeng et al. Silhouette-based gait recognition via deterministic learning
CN107462258B (zh) 一种基于手机三轴加速度传感器的计步方法
Jensen et al. Classification of kinematic swimming data with emphasis on resource consumption
Hasan et al. Robust pose-based human fall detection using recurrent neural network
Geng Research on athlete’s action recognition based on acceleration sensor and deep learning
Lu et al. MFE-HAR: multiscale feature engineering for human activity recognition using wearable sensors
Yuan et al. Adaptive recognition of motion posture in sports video based on evolution equation
CN114881079A (zh) 面向穿戴式传感器的人体运动意图异常检测方法及系统
Javeed et al. Deep activity recognition based on patterns discovery for healthcare monitoring
Ghobadi et al. A robust automatic gait monitoring approach using a single IMU for home-based applications
Hajjej et al. Deep human motion detection and multi-features analysis for smart healthcare learning tools
Sowmyayani et al. Fall detection in elderly care system based on group of pictures
Khandnor et al. A survey of activity recognition process using inertial sensors and smartphone sensors
Bansal et al. Elderly people fall detection system using skeleton tracking and recognition
CN111767932B (zh) 动作判定方法及装置、计算机设备及计算机可读存储介质
CN116092193A (zh) 一种基于人体运动状态识别的行人航迹推算方法
Dorofeev et al. Extraction of individual gait features in mobile phone accelerometer
Sugimoto et al. Robust rule-based method for human activity recognition
Maldonado-Mendez et al. Fall detection using features extracted from skeletal joints and SVM: Preliminary results
Ismael et al. A study on human activity recognition using smartphone

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: 20899467

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: 20899467

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 20899467

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 15/02/2023)

122 Ep: pct application non-entry in european phase

Ref document number: 20899467

Country of ref document: EP

Kind code of ref document: A1