Summary of the invention
The present invention provides a kind of feature extracting method for human body movement data, comprising the following steps:
1) signal data generated when acquiring human motion；
2) the contrary opinion information in the signal data is extracted, the contrary opinion information includes muscle external movement trajectory signal numberIt is believed that vibration signal data information inside breath and muscle；
3) for the contrary opinion information carry out feature extraction, the feature include muscle external movement trajectory signal feature,Vibration signal characteristics and assemblage characteristic, the assemblage characteristic refer to according to the muscle external movement trajectory signal inside muscleThe linked character of vibration signal data information acquisition inside data information and the muscle.
Preferably, the step 1) further comprises:
11) signal data generated when human motion is acquired using single source acceleration transducer；
12) judge the validity of the signal data, and obtain effective signal data
Preferably, the step 2) further comprises:
Low-pass filtering and data cleansing are executed for the signal data, obtains muscle external movement trajectory signal data letterBreath；
Bandpass filtering, data cleansing and personalized difference are executed for the signal data to eliminate, and obtain muscle internal railsMark signal data information；
Wherein, the frequency of the low-pass filtering fluctuates in 10-20 hertz of ranges；The low-limit frequency of the bandpass filtering existsFluctuation, the highest frequency of the bandpass filtering fluctuate in 100-150 hertz of ranges in 5-15 hertz of ranges.
Preferably, for the valid data through the bandpass filtering and data cleansing, maximum spontaneous contractions value is utilizedThe personalized difference is carried out to eliminate, comprising the following steps:
Using range of motion as standard movement, movement number when at least one set of user carries out the standard movement is acquiredAccording to；
Bandpass filtering is executed at least one set of exercise data and data cleansing obtains at least one set of effective exercise numberAccording to；
By the average value of the maximum value of maximum value or multiple groups valid data in the effective exercise data of acquisition, as mostBig spontaneous contractions value；
Normalized is executed using the maximum spontaneous contractions value.
Preferably, the formula for executing normalized using the maximum spontaneous contractions value is as follows:
Wherein, a indicates vibration signal data inside the muscle eliminated by difference, ahinitIndicate first after data cleansingVibration signal inside beginning muscle, MVCperIt is maximum spontaneous contractions value.
Preferably, the step 3) further comprises:
31) it is directed to the contrary opinion information, extracts the muscle external movement trajectory signal feature, muscle internal vibration respectivelySignal characteristic；
32) it is directed to the contrary opinion information, is extracted for describing the muscle external movement track data and the intramuscularThe assemblage characteristic of inner link between portion's vibration data；
33) final feature is obtained using the feature that the step 31) and the step 32) obtain.
Preferably, the assemblage characteristic is that muscle raises energy coefficient feature, and the step 32) further comprises:
321) vibration signal data inside the muscle external movement trajectory signal data and the muscle is sought respectivelyResultant acceleration signal；
322) integral operation is carried out for the resultant acceleration signal obtained, obtains and closes velocity amplitude；
323) muscle is calculated according to following formula raise energy coefficient feature:
Wherein, MREC is that muscle raises energy coefficient feature, for indicating under unit muscle external movement distance, with muscleInside generates and the energy value positive correlation of storage, sh2It is to indicate and the energy generated and store lasting during muscular movementMagnitude positive correlation, slIt is the external movement distance for indicating muscle, vlAnd vhRespectively muscle external movement trajectory signal andThe conjunction velocity amplitude of vibration signal inside muscle.
Preferably, the step 31) further comprises:
311) feature extraction, including average value, standard deviation, frequency are executed for the muscle external movement trajectory signal dataDomain energy, frequency domain entropy；
312) feature extraction is executed for vibration signal data inside the muscle, including cosine correlation, standard deviation, flatEqual power-frequency, power spectrum density, frequency domain entropy.
Preferably, the step 33) further comprises: the characteristic that the step 31) and the step 32) are obtainedSerial end to end splicing is executed, or executes dimension after splicing and about subtracts, obtains final feature.
According to another aspect of the present invention, a kind of human motion identification model method for building up, including following step are also providedIt is rapid:
The characteristic for establishing identification model is extracted using feature extracting method as described above；
Human motion identification model is established using the characteristic.
According to another aspect of the present invention, a kind of human motion identification method is also provided, comprising the following steps:
The motion characteristic data of object to be identified is extracted using feature extracting method as described above；
The motion characteristic data is input to progress human motion identification in human motion identification model.
According to another aspect of the present invention, a kind of feature deriving means for human body movement data are also provided, including withIn the acquisition unit of exercise data acquisition, the feature for extracting the information extraction unit of contrary opinion information and for feature extraction is mentionedTake unit；Wherein, the contrary opinion information extraction unit includes the first letter for extracting muscle external movement trajectory signal dataCease extraction module and the second information extraction modules for extracting vibration signal data inside muscle；The feature extraction unit packetInclude fisrt feature extraction module for extracting muscle external movement trajectory signal feature respectively, for extracting muscle internal vibrationThe second feature extraction module of signal characteristic and the muscle external movement trajectory signal data and described are described for extractingThe third feature extraction module of the assemblage characteristic of inner link between the vibration signal data of muscle inside；The feature extraction unitIt further include for for vibration signal characteristics inside the muscle external movement trajectory signal feature, the muscle and described groupFeature is closed to be combined processing and obtain the processing module of final feature.
According to another aspect of the present invention, a kind of human body movement recognition system, including feature as described above are also providedExtraction element, and identified with data acquisition reception device, model foundation and the movement of feature deriving means communication connectionDevice and recognition result output device.
Compared with the existing technology, the present invention achieves following advantageous effects: human motion identification provided by the inventionSystem and method, by extracting the muscle external movement trajectory signal feature of user, inside muscle vibration signal characteristics and onThe linked character of two kinds of signals is stated, so that obtaining the final feature comprising above-mentioned three kinds of features carries out movement identification, is significantly mentionedHigh accuracy of identification；Particularly, it inventors herein proposes muscle and raises this concept of energy coefficient, can effectively represent above-mentionedThe internal relations of two kinds of signals；In addition, the present invention has only used the acceleration transducer in single source, nothing when carrying out data acquisitionIt needs other sensors to be used cooperatively, not only saves the cost of acquisition equipment, further reduce the wearing weight bearing of user, improve useFamily experience sense is conducive to promote the use of.
Specific embodiment
In order to which the purpose of the present invention, technical solution and advantage is more clearly understood, below in conjunction with attached drawing, to according to thisWhat is provided in the embodiment of invention is further detailed for the feature extracting method and system of human motion, recognition methods and systemExplanation.
The identification of movement non-blood for one group, during exercise due to human body, trunk inherently show very big difference, becauseThis may recognize that this movement over the ground macroscopically using general acceleration transducer；And for the knowledge of one group of relationship movementNot, during exercise due to human body, the overall performance on trunk is similar, therefore also needs to be distinguish by physiological signal.
For this purpose, inventor, by a large amount of experiment discovery, human body during exercise, can not only generate motion profile over the ground,The vibration inside muscle can be also generated simultaneously.It, can be directly using setting in muscle surface for the displacement that opposite ground generatesAcceleration transducer identified to identify motion profile signal (i.e. muscle external movement trajectory signal) over the ground, the letterNumber main band limits be 0~20 hertz, most high frequency can fluctuate within the scope of 15~25 hertz；In addition, since human body existsWhen movement, meat fiber can generate contraction or loosen, so that it is inherent to occur to rub and slide etc. between meat fiberMechanical oscillation, this vibration can reach skin surface by transmitting, and can generate displacement in skin surface, therefore can lead toMeat fiber raises reaction (i.e. muscle inside vibration signal) of the situation in skin surface to human body when crossing acquisition human motionMovement identified, the main band limits of the signal is 5~100 hertz, low-limit frequency can within the scope of 5~15 hertz waveDynamic, highest frequency can fluctuate within the scope of 100~150 hertz.
By the studies above, inventors herein proposes and a kind of realized using acceleration transducer while extracting human body during exerciseVibration signal carries out the system and method for human motion identification inside the muscle external movement trajectory signal and muscle of generation.
Fig. 1 is the human motion identification method schematic diagram that the preferred embodiment of the present invention provides, as shown in Figure 1, with barbell shoulderAway from it is curved lift, barbell width it is curved lift, barbell narrow space is curved lifts, for standard push-up and oblique push-up this five kinds of different movements,Human motion identification method provided by the invention is described in detail.Wherein, barbell shoulder is away from curved act, the curved act of barbell width, thick stickThe curved act of bell narrow space belongs to one group of relationship movement, and standard push-up and oblique push-up belong to one group of relationship and act, and barbell exerciseBelong to non-blood movement between push-up campaign.Human body recognition method provided by the invention the following steps are included:
S10 acquires the signal data that human motion generates
Acquisition device with acceleration transducer is worn on the body of user, specifically, being had using an outsidePocket, and inside is equipped with the movement arm band of acceleration transducer node, and the left arm upper arm of user will be tightly attached on the inside of the movement arm bandAt the center skin of triceps, while it will be arranged in pocket comprising the acquisition device of single-chip microcontroller, Bluetooth communication modules and battery.After the completion of wearing, user can successively carry out one group of above-mentioned five kinds of movement, and acquisition device will record human body and carry out above-mentioned five kinds of fortuneDynamic data.It is accurate in order to calculate, user can be allowed to carry out multiple groups movement respectively.For example, 3 groups of movements can be carried out, every group is completedAfter the completion of movement, user can rest a period of time, then carry out next group of movement.After the completion of acquisition, the data of record can be passed throughBluetooth communication modules in acquisition equipment are transmitted to background server, to execute calculating.
S20 judges data validity
After background server obtains the data of acquisition, will data be carried out with the pretreatment of Effective judgement.Pretreatment is mainIt is to verify the abnormal data for whether containing due to various reasons in collected initial data and leading to occur.It is missed with hardwareIt, can if occurred in the hardware debugging stage similar to hardware problems such as line disconnections for abnormal data caused by poor problemThe data caused are all steady state value, therefore, when being pre-processed, can be used and judge to acquire whether data are that steady state value isThe Effective judgement of preset condition progress data.For example, the data window of a certain section of very little of random selection can be passed through when judgingMouthful, if data value all in this window be all it is identical, illustrate that the data are invalid, and hardware occurProblem needs repairing and acquires equipment；Conversely, then illustrating that the data are effectively, can to continue to execute subsequent step.Especially, in addition to failures such as hardware disconnections mentioned above, data exception caused by other problems is also had, for example, since hardware is setThe problems such as standby electromagnetic interference, lead to abnormal data for occurring fluctuating suddenly etc..
S30 extracts contrary opinion information
Complete above-mentioned steps judgement after, by the valid data of acquisition pass through respectively low-pass filter and bandpass filter intoThe fractionation and data cleansing of row contrary opinion information, to obtain vibration signal inside muscle external movement trajectory signal data and muscleData.Specifically, the frequency of low-pass filter can fluctuate within the scope of 10 hertz to 20 hertz, preferably 15 hertz；Band logicalThe low-limit frequency of filter can fluctuate within the scope of 5~15 hertz, highest frequency can within the scope of 100~150 hertz waveIt is dynamic.After the completion of filtering, data cleansing can be carried out respectively to the data of acquisition, for example, carrying out missing values to obtained filtering signalCompletion, abnormal point removal etc..
Particularly, since different people is when carrying out same action, the average amplitude of vibration signal is usually inside muscleDifferent (for example, the ordinary people seldom moved and fitness enthusiasts), it is therefore, above-mentioned clear by bandpass filtering and data obtainingAfter data after washing, it is also necessary to carry out personalized difference to the data and eliminate, the flesh that can be used to identify human motion could be obtainedMeat internal vibration signal data.Carrying out personalized difference elimination, there are many modes, for example, according to the age of user, height, weightOr exercise habit setting threshold parameter carries out difference elimination.Below to utilize maximum spontaneous contractions value to by bandpass filtering sum numberIt is normalized according to the data after cleaning, to be illustrated for carrying out the mode of personalized difference elimination.
Inventor it has been investigated that, for different people, the average amplitude of vibration signal is usually not inside muscleWith, but for the same person, when carrying out one group of relationship movement, vibration signal amplitude becomes relatively inside muscleGesture is identical.For example, the shoulder for being directed to barbell is moved away from two kinds of curved act and the curved act of narrow space, a fitness enthusiasts are above-mentioned in progressWhen two movements, the amplitude of the internal vibration signal of the bicipital muscle of arm is all larger than the ordinary people that another one seldom moves.But work asFor the two in the curved act of progress narrow space, the muscle vibration signal amplitude of the bicipital muscle of arm but all meets corresponding width when being greater than shoulder away from curved actIt is worth this condition.
Based on this principle, inventor proposes that muscle inside vibration signal data can be carried out MVC (maximum autonomous receiptsContracting) normalized, so as to eliminate because individual habit of user it is different caused by difference in data.Specifically include following stepIt is rapid:
S301, user execute predetermined movement after wearing acquisition device, for example, above-mentioned barbell shoulder is away from curved act, barbell widthCurved act, the curved act of barbell narrow space, standard push-up and oblique push-up, acquisition device can obtain one or more groups of acquisition data；
S302, one or more groups of acquisition data of acquisition are filtered using 5~100 hertz of bandpass filterAnd data cleansing, for example, missing values are not complete and abnormal point removes；
S303, it is maximized for acquisition data after cleaning as maximum spontaneous contractions value MVCper, if it is multiple groups numberAccording to the average value of every group of maximum value can be taken, and using the average value as the maximum spontaneous contractions value for corresponding to the userMVCper。
S304, personalized difference are eliminated
Vibration signal data inside the muscle eliminated by difference is obtained using following formula.
Wherein, a indicates vibration signal data inside the muscle eliminated by difference, ahinitIndicate first after data cleansingVibration signal inside beginning muscle, MVCperIt is everyone maximum spontaneous contractions value.
S40 feature extraction
After completing step S30, the muscle external movement trajectory signal data and flesh that can be used for extracting feature can be obtained respectivelyMeat internal vibration signal data.Feature extraction specifically includes the following steps:
S401, information are divided and ruled processing
For above two signal data, feature extraction is carried out respectively.For example, can using sliding window method respectively fromThe multiple angles such as time domain, frequency domain and time-frequency domain extract.Specifically, muscle external movement trajectory signal data are directed to, it canThe feature of extraction includes but is not limited to average value, standard deviation, frequency domain energy, frequency domain entropy etc., wherein each feature can be from threeA quadrature axis extracts respectively, to obtain 12 dimensional features；For muscle inside vibration signal data, extractible feature include butIt is not limited to cosine correlation, standard deviation, frequency of average power, power spectrum density, frequency domain entropy etc., wherein each feature is okIt is extracted respectively from three quadrature axis, to obtain 15 dimensional features.
S402, information assemblage characteristic extract
The information of completion step S401 is divided and ruled after processing, can obtain the muscle surface of description single user's movement respectivelyWith muscle internal feature, in order to preferably make the two associated, inventor has been researched and proposed a kind of utilization assemblage characteristic descriptionMode when user personality moves outside muscle with muscle internal state, this assemblage characteristic can be by using outside above-mentioned muscleFeature calculation obtains inside portion's feature and muscle.
For this purpose, inventors herein proposing the concept of " muscle raises energy coefficient feature ", table is used for as a kind of assemblage characteristicThe connection between internal vibration signal when showing human motion when the external movement trajectory signal and contraction of muscle of its muscle.
The it is proposed of this concept is the following research based on inventor, since the contraction of human motion and meat fiber hasIt closes, and muscle fibre is during contraction, ATP decomposition makes to store a large amount of elastic potential energy inside muscle.Contraction of muscle is more filledPoint, the value of elastic potential energy is also bigger.Also, when mutual inversion of phases occurs for elastic potential energy release Shi Huiyu kinetic energy, to help fleshMeat generates the displacement relative to ground, that is, forms movement.The process is similar with the movement of spring, and spring is in outer masterpieceElastic potential energy can be generated by being stretched or being compressed with lower generation, and when external force releases, elastic potential energy and kinetic energy can be converted.CauseThis, the kinetic model based on spring, i.e. the energy of spring be it is square directly proportional to distance, the similar dynamic of muscle can be obtainedMechanical model, i.e. elastic potential energy inside muscle are also square directly proportional to distance.But since the elastic potential energy of spring is oneWhat secondary property generated, and the elastic potential energy of muscle is persistently to consume ATP during muscular movement persistently to generate.Therefore, in order toThe kinetic model for improving muscle inventors herein proposes and raises energy coefficient feature about muscle, for qualitatively approximate reflectionMuscle is under unit external move distance out, the value of the internal elastic potential energy for persistently generating and storing, i.e., this muscle bulletProperty potential energy value and muscle raise energy coefficient feature between have positively related relationship.It will be described in detail below and raise energy about muscleThe calculation of coefficient of discharge feature.
Firstly, by vibration signal data point inside the step S30 muscle external movement trajectory signal data obtained and muscleResultant acceleration signal is not sought:
Wherein, alIt is the conjunction signal data of muscle external movement track, alx,aly,alzRespectively indicate muscle external movement railSignal numerical value of the mark signal data on three quadrature axis x, y, z；ahIt is the conjunction signal of muscle internal vibration, ahx,ahy,ahzPointIt Biao Shi not signal numerical value of the muscle inside vibration signal on three quadrature axis.
Secondly, carrying out integral operation to two kinds of conjunction signal datas, obtain closing velocity amplitude:
Wherein, vlAnd vhThe conjunction velocity amplitude of vibration signal respectively inside muscle external movement trajectory signal and muscle.
Finally, executing the calculating that muscle raises energy coefficient feature.
Wherein, MREC is that muscle raises energy coefficient feature, for indicating under unit muscle external movement distance, with muscleInside generates and the energy value positive correlation of storage, sh2It is to indicate and the energy generated and store lasting during muscular movementMagnitude positive correlation, slIt is the external movement distance for indicating muscle.
S50, characteristic processing and movement identification
S501, characteristic processing
The feature that above-mentioned steps S401 and S402 are obtained carries out Fusion Features, for example, serial fusion, Parallel Fusion etc.,To obtain the muscle external movement track characteristic that can be described user's individual during exercise, muscle inside vibration performance and fleshMeat raises energy feature.For example, the end to end acquisition 28 that connects together of the features described above of acquisition can be tieed up using serial modeThe feature of degree, so as to make full use of the characteristic information of all acquisitions.Particularly, in order to prevent because characteristic dimension is excessive andOver-fitting is caused, dimension can also be carried out to the feature of above-mentioned acquisition and about subtracted.
S502, movement identification
After completing above-mentioned steps, disaggregated model is established using the final feature of above-mentioned acquisition and disaggregated model is establishedAfter the completion, the identification of human motion is carried out.Wherein, it is common to can be machine learning for the above-mentioned disaggregated model for moving identificationDisaggregated model.For example, random forest disaggregated model etc..
According to another aspect of the present invention, a kind of system for human motion identification is also provided.Fig. 2 is that the present invention is excellentThe human body movement recognition system structural schematic diagram for selecting embodiment to provide, as shown in Fig. 2, human motion provided by the invention identification systemSystem includes for acquiring the acquisition unit 101 of human motion signal, the information extraction unit 102 for extracting contrary opinion information, usesIt divides and rules processing unit 103, the information combined treatment list for extracting combined information feature in the information for extracting contrary opinion information characteristicsMember 104 and the movement identification model for establishing identification model establish unit 105.
Wherein, signal acquisition unit 101 includes being arranged at muscle to be measured, is close to the acceleration sensing that skin surface is wornDevice.Preferably, the sample frequency of acceleration transducer is not less than 300HZ, and sensitivity is not less than 5000LSB/g, can be conducive toEffective extraction of contrary opinion signal in information extraction unit；Signal acquisition unit 101, can be by acquisition data transmission after the completion of acquisitionTo information extraction unit 102.
Information extraction unit 102 includes for receiving data from the acceleration transducer of above-mentioned signal acquisition unit 101Receiving module and preprocessing module for carrying out data validity judgement to the pending data that receives, utilize pretreatment mouldBlock is completed to can get valid data to after the judgement of received acquisition data validity；Information extraction unit 102 further includes low pass filteredWave module and bandpass filtering modules block, and the low-pass data cleaning module being connected with above-mentioned low-pass filtering module, with above-mentioned band logicalThe connected band logical data cleansing module of filter module and personalized difference cancellation module.Wherein, valid data are low-pass filteredAfter module and low-pass data cleaning module, it can get muscle external movement trajectory signal data, and the data are transmitted to informationIt divides and rules processing unit 103；Valid data are eliminated into mould through bandpass filtering modules block, band logical data cleansing module and personalized differenceAfter block, vibration signal data inside muscle can get, and the data are transmitted to information and are divided and ruled processing unit 103.
Information processing unit 103 of dividing and ruling includes characteristic extracting module inside muscle surface extraction module and muscle,In, muscle surface extraction module is used to carry out feature extraction for the muscle external movement trajectory signal data obtained；FleshMeat internal feature extraction module is used to carry out feature extraction for vibration signal data inside the muscle obtained.
Information combined treatment unit 104 is used for single source contrary opinion information characteristics according to acquisition (for example, muscle external movement railVibration signal characteristics inside mark signal characteristic and muscle) inner link of two kinds of obtained contrary opinion information is extracted, and extract accordinglyThe assemblage characteristic (for example, muscle raises energy coefficient feature) of the two out.
Movement identification model establishes unit 105 and is used for information is divided and ruled processing unit 103 and information combined treatment unit 104It obtains all features and carries out splicing, and identification model is established based on machine learning.
Although in the above-described embodiments, come for the method that personalized difference is eliminated using maximum spontaneous contractions valueIllustrate human motion identification method and system provided by the invention, but it will be recognized by one of ordinary skill in the art that can also useOther methods carry out personalized difference for vibration data inside muscle and eliminate, such as according to age, the exercise habit etc. of user；Although also, in the above-described embodiments, using, muscle raises energy coefficient to indicate muscle external movement trajectory signal and fleshThe inner link of meat internal vibration signal, but it will be recognized by one of ordinary skill in the art that can also be using other way come tableShow connection between the two, such as the weighting scheme based on mathematical principle etc..
Compared with the existing technology, human body movement recognition system provided by the invention and method, using the acceleration in single sourceSensor, equipment cost is low, and acquisition mode is easy, and user experience is more preferable；Meanwhile in feature extraction, using the flesh of userMeat external signal, muscle internal signal and combination signal improve accuracy of identification, are not only applicable in as final featureIn the identification of non-blood movement, additionally it is possible to realize the identification moved to relationship.
Although the present invention has been described by means of preferred embodiments, the present invention is not limited to described hereEmbodiment, without departing from the present invention further include made various changes and variation.