CN105760819B - A kind of daily routines recognition methods based on acceleration signal - Google Patents
A kind of daily routines recognition methods based on acceleration signal Download PDFInfo
- Publication number
- CN105760819B CN105760819B CN201610060643.4A CN201610060643A CN105760819B CN 105760819 B CN105760819 B CN 105760819B CN 201610060643 A CN201610060643 A CN 201610060643A CN 105760819 B CN105760819 B CN 105760819B
- Authority
- CN
- China
- Prior art keywords
- axis
- acceleration
- deflection angle
- signal
- calculating
- Prior art date
- Legal status (The legal status 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 status listed.)
- Expired - Fee Related
Links
Classifications
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
- G06V40/23—Recognition of whole body movements, e.g. for sport training
Abstract
The daily routines recognition methods based on acceleration signal that the invention discloses a kind of; acquisition including acceleration signal; the separation of gravitational acceleration component and movable component of acceleration, the calculating of activity intensity, the calculating of signal low frequency energy and the calculating of acceleration axis deflection angle.Activity intensity is to discriminate between the important indicator of user's static state and dynamic moving, and the low frequency energy of acceleration signal can further discriminate between the type of dynamic moving, and acceleration axis deflection angle helps to judge human body attitude when static state.By features described above combined training multi-categorizer, it can recognize eight kinds of Activity Types common in life, be respectively to walk, go downstairs, go upstairs, run, jump, sit, stand and lie.
Description
Technical field
The daily routines recognition methods based on acceleration signal that the present invention relates to a kind of.
Background technique
The daily routines of user are tracked and identified, can preferably reflect that the behavioural habits of user and daily energy disappear
Consumption has important directive significance to health monitoring, the Aged Care and user situation perception etc..Currently based on wearable sensing
The user behavior recognition of equipment generallys use acceleration and gyroscope signal, and number of features is more, and it is larger to calculate energy consumption.
Therefore, the problem to be solved in the present invention is: being based on 3-axis acceleration signal, one group of proposition is quick to User Activity classification
The feature of sense combines, and preferable User Activity accuracy of identification is realized under low-dimensional feature space, to reduce system energy consumption.
Summary of the invention
To solve the above-mentioned problems, the daily routines recognition methods based on acceleration signal that the present invention provides a kind of is led to
The combination of five dimensional signal features is crossed, identification user walks, goes upstairs, goes downstairs, runs, jumps, sits, stands and lie eight kinds of daily routines.
The method includes signal acquisition, signal characteristic combinations to calculate, multi-categorizer training and daily routines identify four component parts.
Acceleration signal is collected in three axis direction of x, y, z of the positive right side of face hip of human body, and wherein x-axis is vertical direction, y-axis
For left and right directions, z-axis is front-rear direction.
Acceleration signal is divided into the data slot of 5 seconds durations, and signal characteristic calculates the data slice based on 5 seconds durations
Section.
Signal characteristic combination includes activity intensity, x-axis low frequency energy, x-axis deflection angle, y-axis deflection angle and z-axis deflection
Five dimensional feature of angle.
Before calculating each dimensional feature, need to separate the movable component of acceleration and gravity of x, y, z 3-axis acceleration signal
Component of acceleration.Wherein, movable component of acceleration is for calculating activity intensity and axis low frequency energy;Gravitational acceleration component is used for
Calculate axis deflection angle.
Steps are as follows for movable component of acceleration and gravitational acceleration component in separate acceleration signal:
Step 101: 3 median filterings being carried out to acceleration signal, eliminate high-frequency noise;
Step 102: acceleration signal is filtered using the oval IIR low-pass filter that cutoff frequency is 0.25Hz,
Filtering obtained signal is gravitational acceleration component;
Step 103: movable component of acceleration is the difference of acceleration signal and gravitational acceleration component.
Steps are as follows for the calculating of signal characteristic activity intensity:
Step 201: the movable component of acceleration of y-axis and z-axis is calculated, is denoted as y respectivelyAAnd zA;
Step 202: activity intensity
Steps are as follows for the calculating of signal characteristic x-axis low frequency energy:
Step 301: the movable component of acceleration of x-axis is calculated, is denoted as xA;
Step 302: removal xAMean value, i.e. x 'A=xA-mean(xA);
Step 303: to x 'ADiscrete Fourier Transform is sought, spectrum intensity coefficient is obtained;
Step 304: summing to 5Hz spectrum intensity coefficient below, obtain x-axis low frequency energy.
Steps are as follows for the calculating of signal characteristic x-axis deflection angle, y-axis deflection angle and z-axis deflection angle:
Step 401: the gravitational acceleration component of x-axis, y-axis and z-axis is calculated, is denoted as x respectivelyG、yGAnd zG;
Step 402: respectively to xG、yGAnd zGIt negates cosine, obtains angle sequence θx、θyAnd θz。
Step 403: taking θ respectivelyx、θyAnd θzIntermediate value, obtain the deflection angle of x-axis, y-axis and z-axis.That is tiltx=
median(acos(xG)), tilty=median (acos (yG)) and tiltz=median (acos (zG))。
Multi-categorizer training step is as follows:
Step 501: acquisition human body is when walking, going downstairs, go upstairs, run, jump, sit, stand and lie eight kinds of daily routines
3-axis acceleration signal;
Step 502: it uses length 5 seconds, acceleration signal is divided into the data slot of 5 seconds durations by the time window of overlapping 4 seconds,
Establish data set;
Step 503: calculating the five dimensional features combination of data set, including activity intensity, x-axis low frequency energy, x-axis deflection angle
Degree, y-axis deflection angle and z-axis deflection angle, establish training characteristics collection;
Step 504: every one-dimensional characteristic that training characteristics are concentrated is normalized to 0~1 section.
Step 505: using normalized training characteristics collection training multi-categorizer.
The identification step of daily routines is as follows:
Step 601: acquisition human body adds three shaft speed signals;
Step 602: using length 5 seconds, the time window data intercept segment of overlapping 4 seconds;
Step 603: calculating the five dimensional features combination of data slot, including activity intensity, x-axis low frequency energy, x-axis deflection angle
Degree, y-axis deflection angle and z-axis deflection angle.
Step 604: feature combination being normalized using training characteristics collection normalized parameter.
Step 605: corresponding to the class of activity using trained multi-categorizer identification data slot.
The beneficial effects of the present invention are: acceleration signal is used only and identifies daily routines, signal is easily obtained;It proposes
Five dimensional features combine (activity intensity, x-axis low frequency energy, x-axis deflection angle, y-axis deflection angle and z-axis deflection angle) to eight kinds
Daily routines classification (walk, go upstairs, go downstairs, run, jump, sit, stand and lie) has good discrimination, and calculates easy.
Activity intensity is to discriminate between the important indicator of user's static state and dynamic moving, and x-axis low frequency energy can further discriminate between dynamic moving
Type, x, y and z axes deflection angle are beneficial to judge human body attitude when static state.Experimental result shows, one kind proposed by the present invention
Daily routines recognition methods recognition effect based on acceleration signal is good, and the user's daily routines being very suitable under mobile environment are known
Not.
Detailed description of the invention
Fig. 1 is inventive structure schematic diagram;
Fig. 2 is characterized the mean value combined under eight kinds of active states;
Fig. 3 is the acceleration signal schematic diagram of acquisition;
Fig. 4 is the true tag schematic diagram of data;
Fig. 5 is the sequence label schematic diagram of identification.
Specific embodiment
The present invention is described further in the following with reference to the drawings and specific embodiments:
As shown in Figure 1, the present invention needs to establish training dataset before identifying user's daily routines, proposed using the present invention
Five dimensional feature combined training multi-categorizers, dotted line flow elements.After the completion of multi-categorizer training, i.e., according to the spy of data slot
It levies mix vector and carries out Activity Type identification, solid line flow elements.
In order to establish training dataset, acquires 14 subjects and walking, going downstairs, going upstairs, running, jumping, sitting, standing and lying
3-axis acceleration signal when eight kinds of daily routines.It is MotionNode that acceleration, which acquires equipment, and sample frequency 100Hz is adopted
Collection range is ± 6g.
With length 5 seconds, acceleration signal was divided into the data slot of 5 seconds durations by the time window of overlapping 4 seconds, established data
Collection.Comprising walking 3534, sample in data set, goes upstairs 1694, sample, go downstairs 1838, sample, sample 1485 of running
Item jumps 790, sample, sits 2335, sample, stands 2080, sample, lie 3470, sample.
Calculate the five dimensional features combination of data set, including the deflection of activity intensity, x-axis low frequency energy, x-axis deflection angle, y-axis
Angle and z-axis deflection angle.Every one-dimensional characteristic is normalized to 0~1 section, establishes training characteristics collection.Fig. 2 illustrates five Wei Te
Levy the mean value under eight kinds of active states, it can be seen that activity intensity can effectively distinguish static and dynamic behaviour: walking is gone downstairs
Ladder is suitable with the activity intensity gone upstairs, and the activity intensity highest of running hops it, and the activity intensity sat, stand and lain is very low.x
Axis low frequency energy has further differentiation to act on dynamic moving state, and x-axis deflection angle, y-axis deflection angle and z-axis deflection
Angle value of (sitting, stand and lie) in various static activities is very different.
Using normalized training characteristics collection training multi-categorizer, had trained in this example KNN multi-categorizer identification walking, under
Stair, eight kinds of class of activity of going upstairs, run, jump, sit, stand and lie, classifier Neighbourhood parameter k=10.Intersected using 10- folding and is tested
The obtained multi-categorizer of card training, recognition accuracy are walking 99.35%, go downstairs 95.75%, go upstairs 91.46%, running
99.53%, 95.19% is jumped, 97% is sat, stand 98.99%, lies 100%, comprehensive recognition accuracy is 97.75%.
After obtaining multi-categorizer, the identification process of daily routines is illustrated by taking an identification process as an example, such as Fig. 3 to Fig. 5 institute
Show.The activity process of user is to walk 10 seconds, is stood 10 seconds, is walked 5 seconds, is gone downstairs 15 seconds, is run 10 seconds, is amounted to 50 seconds, and acquisition adds
For speed signal as shown in figure 3, the true tag of data is as shown in figure 4, wherein horizontal axis is the time, the longitudinal axis is class of activity label.
To the acceleration signal of acquisition with 5 seconds durations, the time window of overlapping in 4 seconds is divided into data slot, calculates its activity
Intensity, x-axis low frequency energy, five dimensional feature of x-axis deflection angle, y-axis deflection angle and z-axis deflection angle, and use training characteristics
Feature vector is normalized in collection normalized parameter.Feature vector after normalization is inputted into multi-categorizer, is identified
Active tags.And so on, since acceleration signal the 5th second, active tags that each second is all identified.The label of identification
Sequence is as shown in Figure 5.
As seen from Figure 3, five dimensional feature proposed by the present invention combination has preferable class of activity recognition effect, when with
When family persistently carries out certain class activity, the identification output of multi-categorizer is correct and stablizes.User Activity type conversion when, due to when
Between the window data that intercept still include previous state data, the identification output of multi-categorizer has 1-4 seconds unstable.It is whole next
It says, the daily routines recognition methods proposed by the present invention based on acceleration signal has preferable recognition effect, and the spy used
Sign dimension is few, and feature calculation is simple, is very suitable to the identification of daily routines under mobile environment.
The above description of this invention is illustrative and not restrictive, those skilled in the art understand that wanting in right
Ask it can be carried out within the spirit and scope of restriction it is many modification, variation or it is equivalent, but they fall within it is of the invention
In protection scope.
Claims (1)
1. a kind of daily routines recognition methods based on acceleration signal, which is characterized in that the recognition methods includes that signal is adopted
Collection, signal characteristic calculate, multi-categorizer training and daily routines identify four component parts;
Acceleration signal is collected in three axis direction of x, y, z of the positive right side of face hip of human body, and wherein x-axis is vertical direction, and y-axis is a left side
Right direction, z-axis are front-rear direction;
Acceleration signal is divided into the data slot of 5 seconds durations, and signal characteristic calculates the data slot based on 5 seconds durations;
Signal characteristic combination includes activity intensity, x-axis low frequency energy, x-axis deflection angle, y-axis deflection angle and z-axis deflection angle
Five dimensional features;
Before signal characteristic calculates, need to separate the movable component of acceleration and acceleration of gravity of x, y, z 3-axis acceleration signal
Component;Wherein, movable component of acceleration is for calculating activity intensity and axis low frequency energy;Gravitational acceleration component is for calculating axis
Deflection angle;
Steps are as follows for movable component of acceleration and gravitational acceleration component in separate acceleration signal:
Step 101: 3 median filterings being carried out to acceleration signal, eliminate high-frequency noise;
Step 102: acceleration signal being filtered, the signal filtered is gravitational acceleration component;
Step 103: movable component of acceleration is the difference of acceleration signal and gravitational acceleration component;
Steps are as follows for the calculating of signal characteristic activity intensity:
Step 201: the movable component of acceleration of y-axis and z-axis is calculated, is denoted as y respectivelyAAnd zA;
Step 202: activity intensity
Steps are as follows for the calculating of signal characteristic x-axis low frequency energy:
Step 301: the movable component of acceleration of x-axis is calculated, is denoted as xA;
Step 302: removal xAMean value, i.e. x 'A=xA-mean(xA);
Step 303: to x 'ADiscrete Fourier Transform is sought, spectrum intensity coefficient is obtained;
Step 304: summing to 5Hz spectrum intensity coefficient below, obtain x-axis low frequency energy;
Steps are as follows for the calculating of signal characteristic x-axis deflection angle, y-axis deflection angle and z-axis deflection angle:
Step 401: the gravitational acceleration component of x-axis, y-axis and z-axis is calculated, is denoted as x respectivelyG、yGAnd zG;
Step 402: respectively to xG、yGAnd zGIt negates cosine, obtains angle sequence θx、θyAnd θz;
Step 403: taking θ respectivelyx、θyAnd θzIntermediate value, obtain the deflection angle of x-axis, y-axis and z-axis;That is tiltx=median
(acos(xG)), tilty=median (acos (yG)) and tiltz=median (acos (zG));
Multi-categorizer training step is as follows:
Step 501: acquiring three of human body when walking, going downstairs, go upstairs, run, jump, sit, stand and lie eight kinds of daily routines
Axle acceleration signal;
Step 502: using length 5 seconds, acceleration signal is divided into the data slot of 5 seconds durations by the time window of overlapping 4 seconds, is established
Data set;
Step 503: calculating the five dimensional features combination of data set, including activity intensity, x-axis low frequency energy, x-axis deflection angle, y-axis
Deflection angle and z-axis deflection angle, establish training characteristics collection;
Step 504: every one-dimensional characteristic that training characteristics are concentrated is normalized to 0~1 section;
Step 505: using normalized training characteristics collection training multi-categorizer;
The identification step of daily routines is as follows:
Step 601: acquisition human body adds three shaft speed signals;
Step 602: using length 5 seconds, the time window data intercept segment of overlapping 4 seconds;
Step 603: calculating the five dimensional features combination of data slot, including activity intensity, x-axis low frequency energy, x-axis deflection angle, y
Axis deflection angle and z-axis deflection angle;
Step 604: feature combination being normalized using training characteristics collection normalized parameter;
Step 605: corresponding to the class of activity using trained multi-categorizer identification data slot.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610060643.4A CN105760819B (en) | 2016-01-28 | 2016-01-28 | A kind of daily routines recognition methods based on acceleration signal |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610060643.4A CN105760819B (en) | 2016-01-28 | 2016-01-28 | A kind of daily routines recognition methods based on acceleration signal |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105760819A CN105760819A (en) | 2016-07-13 |
CN105760819B true CN105760819B (en) | 2019-03-26 |
Family
ID=56342702
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610060643.4A Expired - Fee Related CN105760819B (en) | 2016-01-28 | 2016-01-28 | A kind of daily routines recognition methods based on acceleration signal |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105760819B (en) |
Families Citing this family (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107016346A (en) * | 2017-03-09 | 2017-08-04 | 中国科学院计算技术研究所 | gait identification method and system |
CN107092861B (en) * | 2017-03-15 | 2020-11-27 | 华南理工大学 | Lower limb action recognition method based on pressure and acceleration sensor |
CN111596325A (en) * | 2020-05-09 | 2020-08-28 | 北京东方海龙消防科技有限公司 | Positioning search and rescue system |
CN112057066A (en) * | 2020-09-03 | 2020-12-11 | 青岛歌尔智能传感器有限公司 | Heart rate detection method, wearable device and computer storage medium |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103400123A (en) * | 2013-08-21 | 2013-11-20 | 山东师范大学 | Gait type identification method based on three-axis acceleration sensor and neural network |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10215587B2 (en) * | 2012-05-18 | 2019-02-26 | Trx Systems, Inc. | Method for step detection and gait direction estimation |
-
2016
- 2016-01-28 CN CN201610060643.4A patent/CN105760819B/en not_active Expired - Fee Related
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103400123A (en) * | 2013-08-21 | 2013-11-20 | 山东师范大学 | Gait type identification method based on three-axis acceleration sensor and neural network |
Non-Patent Citations (3)
Title |
---|
Human activity data discovery from triaxial accelerometer sensor: Non-supervised learning sensitivity to feature extraction parametrization;Ines P. Machado 等;《Infomation Processing & Management》;20151231;第51卷(第2期);204-214 |
人体运动信息获取及物理活动识别研究;刘蓉;《中国博士学位论文全文数据库 社会科学Ⅱ辑》;20091115(第11期);H134-6 |
基于三轴加速度传感器的人体行为识别研究;王洪斌;《中国优秀硕士学位论文全文数据库 信息科技辑》;20150115(第1期);第10-15页第2.3节,第16-18页第3.1、3.2节 |
Also Published As
Publication number | Publication date |
---|---|
CN105760819A (en) | 2016-07-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105760819B (en) | A kind of daily routines recognition methods based on acceleration signal | |
CN105678222B (en) | A kind of mobile device-based Human bodys' response method | |
CN105561567B (en) | A kind of meter step and motion state apparatus for evaluating | |
Andò et al. | A multisensor data-fusion approach for ADL and fall classification | |
EP3563764B1 (en) | Determining body postures and activities | |
JP5920364B2 (en) | Information processing apparatus, representative waveform generation method, and representative waveform generation program | |
Yoneyama et al. | Accelerometry-based gait analysis and its application to Parkinson's disease assessment—part 1: detection of stride event | |
Khan et al. | Accelerometer's position free human activity recognition using a hierarchical recognition model | |
CN104598880A (en) | Behavior identification method based on fuzzy support vector machine | |
CN111568436B (en) | Gait symmetry evaluation method based on regression rotation angle | |
CN106295675B (en) | A kind of Activity recognition method based on smart phone of high accuracy | |
Gjoreski et al. | Accelerometer data preparation for activity recognition | |
Awais et al. | Physical activity classification using body-worn inertial sensors in a multi-sensor setup | |
Gao et al. | A comparison of classifiers for activity recognition using multiple accelerometer-based sensors | |
Hellmers et al. | Stair climb power measurements via inertial measurement units | |
Wójtowicz et al. | Fall detector using discrete wavelet decomposition and SVM classifier | |
Ghobadi et al. | A robust automatic gait monitoring approach using a single IMU for home-based applications | |
Choi et al. | Gait type classification using smart insole sensors | |
CN103632133A (en) | Human gesture recognition method | |
AlZubi et al. | Human activity classification using a single accelerometer | |
CN111544005B (en) | Parkinson's disease dyskinesia quantification and identification method based on support vector machine | |
Nouredanesh et al. | Detection of compensatory balance responses using wearable electromyography sensors for fall-risk assessment | |
KR102194313B1 (en) | Apparatus and method for identifying individuals by performing neural network analysis for various detection information | |
Zhang et al. | An efficient feature selection method for activity classification | |
Özel et al. | Implementation of Artifact Removal Algorithms in Gait Signals for Diagnosis of Parkinson Disease. |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190326 Termination date: 20210128 |