CN106326906B - Activity recognition method and device - Google Patents

Abstract

Embodiments of the present invention provide a kind of activity recognition method and device.This method comprises: every class movable radio frequency discrimination RFID signal of the acquisition in the predefined activity of multiclass；For every class activity, multiple active characteristics are extracted from the RFID signal；And the multiple active characteristics are learnt, to construct an associated active dictionary for every class activity, so that activity classification to be identified is one of predefined activity of the multiclass by constructed active dictionary.Due to being carried out for the movable dictionary learning of every class and training process independently of other activities, so that activity recognition method and device have flexibility and scalability.

Description

Activity recognition method and device

Technical field

Embodiment of the present invention relate to activity recognition, more particularly, to based on dictionary activity recognition method and Device.

Background technique

Since average life span growth and birth rate decline, world population is just tending to aging.With cheap biography The newly-developed of sensor and network technology, exploitation is widely applied, such as remote health monitoring and intervention have become a kind of possibility. These applications provide the potentiality for improving life of elderly person quality, give them the bigger sense of security, and promote theirs It lives on one's own life.

For example, how completely and one the daily daily life of the people by monitoring with dementia, can track daily daily life It is performed with causing, and can determine when the patient wants help.Centered on recognizing these applications, activity recognition is in recent years Occur as an important research and development field.Mankind's activity identification relevant to computer vision is one of research direction. Unfortunately, scheme relevant to computer vision needs people's monitoring or high-tech to carry out machine translation at original. In addition, the performance of the mode of this view-based access control model depends critically upon the direction of lighting condition and monitoring camera, this is greatly limited Its applicability in actual deployment is made.Importantly, monitoring camera is typically considered to invade the privacy of people, and It is unsuitable for using under privacy contexts, such as is used in house.

With radio frequency identification (radio frequency identification, RFID) technology and wireless sensor network The increasingly maturation of technology becomes according to inertia, the reading of non-invasi sensor to carry out activity recognition in the past few years One popular research field.Inertial sensor is used as the wearable sensor identified for mankind's activity extensively.Although with Traditional mode based on computer vision is compared, and sensor-based activity recognition can better solve problem, such as hidden Private problem, but most of sensor-based activity recognition schemes require people and dress sensor.These wearable biographies Sensor is usually battery powered, main disadvantage is that, need to safeguard that (for example, replacement battery) is cooperated with user.Therefore, this Modes are not always practical a bit, for the elderly particularly with monitoring with cognitive disorder.

Therefore, it is necessary to the activity recognition schemes of a kind of effective, inconspicuous (unobtrusive).

Summary of the invention

In order to overcome the above problem, the invention proposes a kind of disclosed based on the mode of learning of dictionary it is different movable Structural information between RFID signal, to provide a kind of scheme for carrying out activity recognition based on active dictionary.

In the first aspect of the present invention, a kind of activity recognition method is provided.This method comprises: acquisition is pre- from multiclass The movable radio frequency discrimination RFID signal of every class in the activity of definition；For every class activity, mentioned from the RFID signal Take multiple active characteristics；And the multiple active characteristics are learnt, to construct a correlation for every class activity The active dictionary of connection, so that activity classification to be identified is the predefined activity of the multiclass by constructed active dictionary One.

In one embodiment, each active dictionary in constructed active dictionary includes for characterizing respective activity , the set of the basis vector of linear independence.

In one embodiment, this method further comprises: being directed to every class activity, it is special to calculate the multiple activity The correlation between every a pair of of active characteristics in sign；It is selected from the multiple active characteristics based on the calculated correlation The active characteristics of predetermined number；And selected active characteristics are learnt, to construct work associated with respective activity Dynamic dictionary.

In one embodiment, predetermined number is selected from the multiple active characteristics based on the calculated correlation Active characteristics include: sequence according to correlation calculated from low to high, the multiple active characteristics are ranked up；With And the active characteristics of the predetermined number are selected from ranked multiple active characteristics.

It in one embodiment, include: sound for one of predefined activity of the multiclass by activity classification to be identified Ying Yu is received from the movable inquiry RFID signal to be identified, extracts multiple work from the inquiry RFID signal Dynamic feature, to generate query feature vector set；Using the set of the basis vector in each active dictionary come right respectively The query feature vector set carries out sparse coding, to generate the signal of multiple reconstructions；By the query feature vector set It is compared with each of the signal of the multiple reconstruction；And in response to the query feature vector set and described more One of the signal of a reconstruction matches, and is work associated with the signal of matched reconstruction by the activity classification to be identified It is dynamic.

In one embodiment, by each of the signal of the query feature vector set and the multiple reconstruction Being compared includes: reconstruction between each of signal for calculating the query feature vector set and the multiple reconstruction Error；It and include: wherein sound for activity associated with the signal of matched reconstruction by the activity classification to be identified Reconstruction error between one of query feature vector set and the signal of the multiple reconstruction described in Ying Yu is minimum, by described wait know Other activity classification is activity associated with the smallest reconstruction error.

In one embodiment, extracting multiple active characteristics includes extracting multiple statistical natures.

In one embodiment, the multiple statistical nature includes at least three in the following terms: the radio frequency is known The minimum value of other RFID signal；The maximum value of the radio frequency discrimination RFID signal；The mean value of the radio frequency discrimination RFID signal；Institute State the variance of radio frequency discrimination RFID signal；The root mean square of the radio frequency discrimination RFID signal；The radio frequency discrimination RFID signal Standard deviation；And the intermediate value of the radio frequency discrimination RFID signal.

In one embodiment, it acquires from RFID receiver from every movable RFID signal of class；And And it wherein executes and physical contact is not present between the movable object of every class and the RFID receiver.

In the first aspect of the present invention, a kind of activity recognition device is provided.The device includes: signal acquisition unit, quilt It is configured to acquire the movable radio frequency discrimination RFID signal of every class in the predefined activity of multiclass；Feature extraction unit, It is configured as extracting multiple active characteristics from the RFID signal for every class activity；And unit, it is configured To learn to the multiple active characteristics, to construct an associated active dictionary for every class activity, so that Activity classification to be identified is one of predefined activity of the multiclass by constructed active dictionary.

In the present invention, active dictionary out is learnt by unsupervised sparse coding algorithm.Know with existing activity Other mode is compared, and the mode of learning of the invention based on dictionary realizes movable more compact expression, while remaining more Add information abundant, thus constitutes the basis of effective and healthy and strong identification of object activity.

Further, since being carried out for the movable dictionary learning of every class and training process independently of other activities, thus make Obtaining activity recognition method and device has flexibility and scalability, because not needing when adding New activity to existing work Dynamic dictionary is changed.

Detailed description of the invention

It refers to the following detailed description in conjunction with the accompanying drawings, the above and other feature, advantage of each embodiment of the present invention and side Face will be apparent.In the accompanying drawings, the same or similar appended drawing reference indicates the same or similar element, in which:

Fig. 1 shows embodiments of the present invention and can be implemented in exemplary environments therein；

Fig. 2 shows from " walking about " movable RFID signal strength fluctuation, the fitting to RFID signal strength fluctuation And corresponding residual error；

Fig. 3 a shows the distribution and accumulated probability for " walking about " movable RSSI；

Fig. 3 b shows the distribution and accumulated probability for " kicking left leg " movable RSSI；

Fig. 4 shows the flow chart of activity recognition method according to embodiment of the present invention；

Fig. 5 schematically depicts activity recognition procedure according to one embodiment of the present invention；

Fig. 6 a shows the correlation between " maximum value " feature in the space 2D and " mean value " feature；

Fig. 6 b shows the correlation between value tag minimum in the space 2D and Variance feature；

Fig. 7a-7c illustrates influence of the parameter to activity recognition；

Fig. 8 shows the performance comparison result of method and existing method of the invention；

Fig. 9 is shown in a manner of confusion matrix comments the detailed performance of the activity recognition method of embodiment according to the present invention Estimate；

Figure 10 a-10d shows between feature selection approach and existing feature selection approach based on CCA, in efficiency and has Comparison result in terms of effect property；And

Figure 11 shows the block diagram of activity recognition device according to one embodiment of the present invention.

Specific embodiment

The embodiment that the present invention will be described in more detail below with reference to accompanying drawings.Although showing certain of the invention in attached drawing A little embodiments, it should be understood that, the present invention can be realized by various forms, and be not construed as limiting In embodiments set forth herein, providing these embodiments on the contrary is in order to more thorough and be fully understood by the present invention.It answers When understanding, being given for example only property of drawings and the embodiments effect of the invention is not intended to limit protection model of the invention It encloses.

Term " includes " as used herein and its deformation are that opening includes, i.e., " including but not limited to ".Term "based" It is " being based at least partially on ".Term " embodiment " expression " at least one embodiment "；Term " another embodiment party Formula " expression " at least one other embodiment ".The related definition of other terms provides in will be described below.

With reference first to Fig. 1, it illustrates embodiments of the present invention can be implemented in exemplary environments 100 therein.In In environment 100, RFID antenna (also referred to as RFID transmitter) 110 is configured as sending radiofrequency signal to RFID receiver 120 And/or radiofrequency signal is received from RFID receiver 120.

RFID receiver 120 include by RFID label tag (tag) 1201 ..., 1204 ..., 1209 ..., 1212 constitute RFID label tag array.RFID label tag can be passive, active or battery assisted passive.Active RFID tag has plate It carries battery and periodically sends its ID signal.The RFID label tag of battery assisted passive has the baby battery on plate and it exists It is activated in the case where RFID reader.Passive RFID tags do not include battery, thus more cheap and volume is smaller.In the present invention Embodiment in, it is preferred to use passive RFID tags.However, using active or battery assisted passive RFID label tag It is possible.The scope of the present invention is not limited in this respect.

RFID antenna 110 and RFID receiver 120 can be for example respectively arranged on indoor wall, so that they can be grasped Make to form RFID signal field, monitored object can the execution activity in being formed by RFID signal field.RFID antenna 110 The radiofrequency signal emitted generates reflection after encountering monitored object, and radiofrequency signal through reflection carries back monitored object Action message.RFID receiver 120 receives radiofrequency signal through reflection.

RFID reader (reader) 130 carries via RFID antenna 110 from the reading of each RFID label tag monitored The radiofrequency signal of the action message of object.In the context of this application, penetrating for the action message of monitored object is carried Frequency signal is also referred to as radio frequency identification (RFID) signal.

It is specific to calculate equipment, such as desktop computer, laptop computer, tablet computer etc., it can be with RFID Reader 130 is communicated, and to obtain RFID signal from it, and then the intensity by analyzing acquired RFID signal is come to right The activity of elephant is identified.

It is appreciated that generalling use received signal strength indicator (Received Signal in the field RFID Strength Indicator, RSSI) characterize the intensity of RFID signal.It is well known that due to there are signal reflex, diffraction and Scattering, thus the RSSI in true environment be it is extremely complex, for passive RFID tags.RSSI usually by The attribute or object of object accompanying by label to communication environments and in signal coverage areas are mobile to be seriously affected.Nothing The signal strength of source RFID label tag is uncertain and is nonlinear.The top half of Fig. 2 shows from 12 RFID and marks Label acquisition from " walking about " movable RFID signal strength fluctuation, the lower half portion of Fig. 2 is shown to RFID signal intensity Fluctuation it is linear, secondary, three times with fitting of a polynomial and corresponding residual error.As shown in Fig. 2, the fluctuation of RSSI be not easy using General linear regression and polynomial regression are fitted, because regression criterion is quite big.Therefore, it is not possible to straight in activity recognition It connects and uses original RSSI data.

In addition, although RFID signal intensity reflects as described above uncertain and nonlinear Distribution mode, I Find that the fluctuation (variation) of RFID signal intensity reflects different activity patterns, so as to according to RFID collected believe Number intensity data distinguishes different activities.Fig. 3 a shows the distribution and accumulated probability for " walking about " movable RSSI, figure 3b shows the distribution and accumulated probability for " kicking left leg " movable RSSI.From Fig. 3 a and 3b as can be seen that from above-mentioned The distribution of two movable RSSI and accumulated probability are significantly different.Thus, we believe that the radiofrequency signal of passive RFID tags is strong Degree fluctuation embodies different movable modes, can use this point to carry out effective activity recognition.

Therefore, in one aspect of the invention, a kind of activity recognition method based on dictionary is provided.Fig. 4 shows root According to the flow chart of the activity recognition method 400 of an embodiment of the invention.As shown in figure 4, in step S410, acquisition is come from The movable radio frequency discrimination RFID signal of every class in the predefined activity of multiclass.In step S420, for every class activity, Multiple active characteristics are extracted from the RFID signal.Then, in step S420, the multiple active characteristics are learnt, To construct an associated active dictionary for every class activity, so that constructed active dictionary is by activity to be identified It is classified as one of predefined activity of the multiclass.

Fig. 5 schematically depicts activity recognition procedure accord to a specific embodiment of that present invention.For collection activity letter Breath predefines (design) posture and movement of the most common following 23 class orientation-sensitive in people's daily routines: (1) sitting Directly, (2) are sat to the left, and (3) are sat to the right, and (4) sit back in one's chair, and (5) sitting posture is antecurvature, and (6) stand erectly, (7) from being seated to standing, (8) it walks about, (9) hold arm high and wave (both hands level is brandished), and (10) hold arm high and wave and (brandish before and after both hands), and (11) are held high Arm waves (one hand brandish to the left), and (12) hold arm high and wave (one hand is brandished to the right), and (13) kick forward left leg, (14) to the left Left leg is kicked, (15) kick left leg backward, and (16) kick forward right leg, and (17) kick to the right right leg, and (18) kick right leg backward, and (19) bend over, (20) from couchant to standing, (21) fall forward, and (22) fall to the left, and (23) fall to the right.

In one embodiment, it is desirable that object executes above-mentioned predefined each posture or acts and holds pre- timing Between, such as 120 seconds.RFID reader 510 is marked via RFID antenna (such as RFID antenna 110 shown in FIG. 1) from each RFID Label (such as RFID label tag shown in FIG. 1) acquisition carries the radiofrequency signal of the action message of object, i.e. RFID signal.

As previously mentioned, the fluctuation of RFID signal intensity reflects different activity patterns.Therefore, in an embodiment In, the strong of the radiofrequency signal for carrying the action message of object is directly acquired from each RFID label tag using RFID reader 510 Degree evidence, the i.e. intensity data of RFID signal, using as the activity data for activity recognition.As an example, RFID believes Number intensity characterized by RSSI.Hereinafter, for ease of description, the processing to activity data is described by taking RSSI as an example. It will be appreciated, however, that with the development of technology, it is also possible that the intensity of RFID signal is characterized using other parameters.

RSSI data collected are stored in database 520.In one embodiment, it will be executed for object complete The predefined posture of above-mentioned 23 class in portion or movement and the RSSI data acquired regard a set of activity data as.By the set It is divided into the training dataset for feature selecting and identification model reconstruction and the test data for being assessed model Collection, as shown in Figure 5.

Due to the form that RSSI data collected are continuous data stream sequences, and continuous data stream sequences include big The information of amount, it is difficult to directly it be handled, therefore usually continuous data stream sequences are split first, to be formed Multiple independent data sectionals (segment), as shown in the step 530 of Fig. 5.Being split to continuous data stream sequences can be with It is carried out according to mode well known by persons skilled in the art.

As an example, will can be divided into a static manner for the movable RSSI data stream sequences of every class has admittedly Multiple data sectionals of measured length.In static mode, each data sectional can have length appropriate, such as each data Segmentation includes 6 samples.

As another example, will can be divided into a dynamic fashion for the movable RSSI data stream sequences of every class has Multiple data sectionals of variable-length.

For purposes of illustration, in the context of this application, it will be described by taking static segmentation as an example.Pass through static state point It cuts, RSSI data stream sequences S is divided into the set of the segmentation of the independent data with regular length.RSSI data flow sequence as a result, Column S can be expressed as S={ S₁,…,S_n, wherein S_iIndicate i-th of data sectional, i=1 ..., n.For example, in data sectional S_i Including 6 samples and in the case where using 12 RFID label tags, data sectional S_iIt can be represented as 12 rows × 6 column square Battle array.

Then, as shown in the step 540 of Fig. 5, active characteristics are extracted from each data sectional.In an embodiment In, it include being mentioned from from the predefined movable each data sectional of every class from active characteristics are extracted in each data sectional Take multiple statistical natures.Table 1 shows the example of multiple statistical natures.

Number	Feature	Description
			1	Minimum value	Data sectional SiMinimum value
2	Maximum value	Data sectional SiMaximum value
			3	Mean value	Data sectional SiMean value
4	Variance	Data sectional SiVariance
			5	Root mean square	Data sectional SiRoot mean square
6	Standard deviation	Data sectional SiStandard deviation
			7	Intermediate value	Data sectional SiIntermediate value

Table 1

It should be appreciated that statistical nature shown by table 1 is illustrative and not restrictive.It can be from each data sectional Extract all or part of above 7 statistical natures.Alternatively, it is possible to be extracted from each data sectional except above 7 systems Count other features except feature.The scope of the present invention is not limited in this respect.

In general, indicating that active characteristics are (hereinafter referred to as " special using active eigenvector (hereinafter referred to as " feature vector ") Sign "), and use active eigenvector set (hereinafter referred to as " feature vector set ") O={ o₁,…,o_mIndicate whole spies Vector is levied, wherein o_iIndicate that ith feature vector, i=1 ..., m, m are the number of the feature vector in feature vector set.In In the case where 7 statistical natures of all of the above, feature vector set O includes 7 feature vectors (m=7).It is appreciated that Feature vector o_iDimension, i.e. feature vector o_iIncluded in component (component) number, depend on used The number of RFID label tag.For example, using 12 RFID label tags, feature vector o_iDimension be 12.

As an example, in data sectional (matrix) S_iInclude the case where 6 samples and uses 12 RFID label tags Under, for matrix S_iIn every a line, the minimum value of all elements in the row is calculated, to obtain and data sectional S_iIt is corresponding Minimum value feature vector o₁, wherein minimum value feature vector o₁Include 12 components, i.e. minimum value feature vector o₁For 12 row × 1 The vector of column.Similarly, matrix S can be directed to_iIn every a line, calculate the maximum value, mean value of all elements, side in the row Difference, root mean square, standard deviation and intermediate value, to obtain corresponding feature vector o₂To o₇.It obtains as a result, and data sectional S_iIt is corresponding Feature vector set O={ o₁,…,o_m, wherein m=7.And so on, it can be for from the movable each data of every class Segmentation carries out above-mentioned calculating, to obtain multiple feature vector set with every class activity association.It is appreciated that obtained spy The number for levying vector set depends on the number of divided data sectional.For example, in the RSSI data from specific activities Flow sequence S={ S₁,…,S_nIn the case where, by above-mentioned calculating, the available n set of eigenvectors with the activity association It closes.Collect as a result, with this n feature vector set of the activity association, i.e. the matrix that n feature vector set is constituted, It is used as the training sample set identified to the activity.

It is well known that high quality, have the feature of distinction to the classification accuracy for improving any pattern recognition system be to It closes important.However, some features may make separator fascination without being to aid in classifier to distinguish activity.Further, since " dimension Spend disaster ", when no enough training datas carry out reliably all parameters of learning activities model, as more features is added Add, classification performance may sharply decline.In order to reach optimal classification performance, the number of feature vector should be as small as possible, To only retain most prominent and complementary feature.

Fig. 6 a shows the correlation between " maximum value " feature in the space 2D and " mean value " feature, and Fig. 6 b is shown Correlation between value tag minimum in the space 2D and Variance feature.As can be seen that although characteristics of mean can be with from Fig. 6 a The activity of " walking about " is roughly distinguished with other four kinds activities, but two features (maximum value and mean value) cannot be well Distinguish five kinds of activities.In figure 6b, Variance feature can help to identify that " hold arm high and wave (level) " activity and " bending over " are living It is dynamic, but intersect and be overlapped due to existing, thus both activities can not be characterized using both Variance feature and minimal characteristic. In this case, minimum value tag is considered unrelated or redundancy, and the information being not provided with is to improve Separate accuracy.

In order to systematically evaluate serviceability and identify for distinguishing different movable most important features, need to use Feature Selection, as shown in the step 550 of Fig. 5.Specifically, it in step 550, for every class activity, calculates extracted The correlation between every a pair of of feature in multiple features, and based on the calculated correlation and from the multiple active characteristics The active characteristics of middle selection predetermined number.

In one embodiment, canonical correlation analysis (canonical correlation analysis, CCA) is utilized To calculate the correlation between every a pair of of feature in extracted multiple features.

Specifically, it enablesIt is living for the kth class in multiple predefined activities It is dynamicTraining sample set, wherein N be from activityThe number of data sectional that is divided into of RSSI data stream sequences S Mesh,For with data sectional S_iCorresponding set of eigenvectors merge andThe dimension that m is characterized.

As previously mentioned, feature vector setIt can be represented asWherein m is feature vector Number, i.e., the number of extracted feature.Given a pair of feature vector o_IAnd o_J, find out corresponding linear combinationWithSo thatWithThere is maximum correlation in the space that dimension reduces, that is, by making following equation It is maximized to obtain canonical correlation coefficient ρ:

Wherein,And

As an example, using 7 statistical natures shown in table 1, feature vector setIncluding 7 A feature vector (m=7).By to every a pair of of feature vector application CCA in 7 feature vectors, available 7 rows × The canonical correlation coefficient matrix of 7 column.It similarly, can be to from activityEach data sectional application CCA, to obtain N A canonical correlation coefficient matrix.Then, mean value is taken to the corresponding element in N number of canonical correlation coefficient matrix, to obtain final Canonical correlation coefficient matrix, referred to as " similarity matrix ".

For example, similarity matrix P can be represented asWherein P_IJFor feature vector o_I And o_jCanonical correlation coefficient mean value and | p_ij|≤1.For example, in feature vector o_iIt indicates and extracted maximum value tag pair The feature vector answered, feature vector o_jIn the case where indicating feature vector corresponding with extracted minimum value tag, p_ijIt indicates The correlation of maximum value tag and minimum value tag.

Canonical correlation coefficient p_ijAbsolute value | p_ij| it is bigger, show that correlation is stronger between feature i and feature j, connection is got over Closely；|p_ij| it is smaller, show that correlation is weaker between feature i and feature j, connection is not close.The weaker feature pair of correlation It is considered feature complimentary to one another, that distinction is strong, and the stronger feature of correlation is superfluous each other to being considered The feature remaining, distinction is weak.Therefore, it in order to improve the classification accuracy of behavior identity system, should select as far as possible each other The feature complementary, distinction is strong.

In one embodiment, feature selecting can be carried out based on similarity matrix obtained.

As an example, entire training sample set is traversed using sweep forward (forward selection) algorithm Close O^k, therefrom to select character subset.In sweep forward, character subset is since empty set, and selection one is seen currently every time Optimal feature is added in character subset being so that character subset becomes larger, when reach scheduled number of features or When all features are all contemplated by, search process is terminated.

Specifically, initially, from training sample set O^kIn be randomly chosen a feature vector, such as o_i, and will be special Levy vector o_iIt is added in character subset.Next, from training sample set O^kOne candidate feature vector of middle reselection, such as o_j.Then, similarity matrix P is searched for, to determine feature vector o_iWith o_jCanonical correlation coefficient absolute value | p_ij| it whether is phase It is the smallest in absolute value like all canonical correlation coefficients in property matrix P.If it is, by feature vector o_jIt is added to feature In subset.Otherwise, not by feature vector o_jIt is added in character subset.Then, continue from training sample set O^kMiddle selection one New candidate feature vector, and above-mentioned search process is executed to similarity matrix P, if the spy not being considered compared to other Levy vector, the canonical correlation coefficient in the new candidate feature vector and character subset between existing each feature vector it is exhausted It is the smallest to value, it is determined that add it in character subset；Otherwise, it does not add it in character subset.It is pre- when reaching When fixed number of features or all features are all contemplated by, search process is terminated.

It is selected back to Fig. 5 by features described above selection course from being predefined in movable training sample set from every class Corresponding character subset is selected out, as shown in step 560.Feature in this feature subset is considered as complimentary to one another, differentiation The strong feature of power can be realized more accurate activity recognition using these features.It should be appreciated that features described above selection course essence On be to calculate correlation between every a pair of of feature in multiple features, according to the calculated each pair of spy of institute for every class activity The sequence of correlation from low to high between sign is to be ranked up multiple features and from ranked multiple active characteristics Select the process of the active characteristics of predetermined number.

After obtaining and predefining movable character subset for every class, our target is to for the predefined work of every class Dynamic character subset (selected feature) is learnt, for predefined activity building one associated movable word of every class Allusion quotation, so that activity classification to be identified is one of predefined activity of multiclass by constructed active dictionary, such as step 570 institute Show.

In one embodiment, each active dictionary in constructed active dictionary includes for characterizing respective activity , the set of the basis vector (basis vector) of linear independence.Each basis vector can be captured effectively/be indicated and come from In the key structure information of the movable given training sample set of every class.Particularly, features described above selection course is being executed to generate In the case where for the movable character subset of every class, each basis vector can effectively capture/indicate the key of character subset Structural information.

In one embodiment, it is obtained by being solved to sparse optimization problem basic in each active dictionary The set of vector.About the process that sparse optimization problem is solved E.J.Candes, J.Romberg and T.Tao with It is described in detail in Publication about Document: " Robust uncertainty principles:Exact signal reconstruction from highly incomplete frequency information"；Information Theory, IEEE Transactions, 52 (2): page 489-509,2006.The side that disclosures of the above documents pass through reference Formula is fully incorporated in this.

Hereinafter, the active dictionary learning process of embodiment according to the present invention will be described in conjunction with specific example.

As previously mentioned, enablingFor from kth class activityTraining sample set It closes, wherein N is from activityThe number of data sectional that is divided into of RSSI data stream sequences S,To divide with data Section S_iCorresponding set of eigenvectors merge andThe dimension that m is characterized.

In order to learn and encode the information for belonging to specific activities class testing sample, first against every class activityBuilding Cross complete (overcomplete) dictionaryWe, which are intended to find one, has K (K > m) a basis vectorDictionary matrixSo that O^kIn dictionary matrixIt is upper that there is sparse table ShowWhereinIt is so-called " sparse " to refer to: X^kIn it is only seldom several Nonzero element or it is only seldom it is several be much larger than zero element.In this case, original training sample set (matrix) O^kIt can be represented as being no more thanThe linear combination of a basis vector.Due to X^kIn it is only seldom several Nonzero element, therefore for an inquiry RFID signal, it is only necessary to less several basis vectors can be by the RFID signal It shows.Optimization problem can be represented as:

Wherein, ‖ ‖₂2 norm operations are sought in expression.

In one embodiment, using K mean value singular value decomposition (K-means Singular Value Decomposition, SVD) algorithm solves above formula (3).In K-SVD algorithm, it is straight to be iteratively performed two steps To convergence.First step is the sparse coding stage, whereinIt is kept fixed, and passes through orthogonal matching pursuit (Orthogonal Matching Pursuit) algorithm calculates the sparse coefficient matrix X of rarefaction representation^k.Then, continuously more New dictionaryAllowing relevant sparse coefficient is unique for K-SVD algorithm, and causes to restrain faster.About utilization The process that K-SVD algorithm solves above formula (3) in the following documents of M.Aharon, M.Elad and A.Bruckstein into Detailed description is gone: " K-svd:An algorithm for designing overcomplete dictionaries for sparse representation"；Signal Processing, IEEE Transactions, 54 (11): 4311-4322 Page, 2006.Disclosures of the above documents are fully incorporated in this by reference.

Pseudo-code for realizing above-mentioned dictionary learning process is as follows:

There are several advantages for embodiment according to the present invention and the active dictionary created.Firstly, movable for every class Active dictionary is learnt from the collecting of training sample, and solving to sparse optimization problem, with more compact Bigger mode indicates the structural informations of RSSI data with information content.Secondly, being directed to the movable dictionary learning of every class and training Process is carried out independently of other activities, this makes behavior identity system have flexibility and scalability, because when addition is new When movable, do not need to be changed existing active dictionary.Finally, each active dictionary can be by being used only a small amount of instruction Practice sample to train and learn, this can effectively alleviate in most of existing activity recognition methods to training data mark-on It signs (label) and marks the heavy burden of (annotate).

It, will be wait know using constructed active dictionary after constructing an associated active dictionary for every class activity Other activity classification is that one of predefined activity of multiclass includes: to be identified movable to look into response to receiving from described RFID signal is ask, multiple active characteristics are extracted from the inquiry RFID signal, to generate query feature vector set；Using institute The set for stating the basis vector in each active dictionary to carry out sparse coding to the query feature vector set respectively, with life At the signal of multiple reconstructions；Each of the query feature vector set and the signal of the multiple reconstruction are compared Compared with；And match in response to one of the query feature vector set and the signal of the multiple reconstruction, it will be described to be identified Activity classification be activity associated with the signal of matched reconstruction.

In one embodiment, by each of the signal of the query feature vector set and the multiple reconstruction Being compared includes: reconstruction between each of signal for calculating the query feature vector set and the multiple reconstruction Error；It and include: wherein sound for activity associated with the signal of matched reconstruction by the activity classification to be identified Reconstruction error between one of query feature vector set and the signal of the multiple reconstruction described in Ying Yu is minimum, by described wait know Other activity classification is activity associated with the smallest reconstruction error.

Specifically, after constructing active dictionary for every class activity, for the given query characteristics of sample of signal Vector set o^*, for k-th of movable reconstruction error e_kIt can calculate as follows:

e_k=| | o^*-D^kX^k||₂(7) wherein, [1, K] k ∈, the number of K expression activity class.

It is then possible to using following formula come for query feature vector set o^*Allocation activities label l_o*:

Pseudo-code for realizing the above-mentioned activity classification process based on signal reconstruction is as follows:

In order to verify the activity recognition scheme of embodiment according to the present invention, many experiments have been carried out.It hereinafter, will be first Key Experiment setting is first briefly introduced, is acquired including hardware setting, sampling rate and data.Then, it illustrates to being proposed Scheme a large amount of experimental research achievements.

Hardware setting.In an experiment, using an Alien 9900+RFID reader, four circular antenna (each antennas For a room) and 12 Squig edge body (inlay) passive RFID tags.RFID label tag is placed on wall, forms 4 × 3 RFID label tag array.Based on the empirical research disposed to label, we roughly set each lattice of the array to 0.8m×0.8m.Antenna is disposed in the height of ≈ 1.3m~1.6m, towards 70 ° of label ≈.It is adopted with the sampling rate of 0.5s Collect RSSI.

One is the problem of meriting attention, and the behavior identity system based on RFID may cause damages to the health of people.Quotient It is operated under the electromagnetic frequency of low energy range with RFID reader and label, effectively eliminates the wind with human cell's interaction Danger.In addition, passive RFID tags itself do not have baseline electromagnetic activity, and it is only in response to the inquiry from RFID reader (interrogation) signal is generated.RFID label tag itself or even it has been allowed to be implanted into human body and has not shown and take the post as What negative health effects.In the present invention, it acquires from RFID receiver from the movable RFID signal of every class, and execution activity Object and the RFID receiver between there is no physical contact.Hereby it is achieved that the activity of equipment unrelated (device-free) Identification.

Sampling rate.Passive RFID tags are often noisy, or even in laboratory environment.For example, existing RFID system One in system challenge is mistake caused by by missing detection (that is, label is in the read range of antenna, but being not detected) Negative readings accidentally.Meanwhile RSSI data are very sensitive to environment, i.e. some interference from environment can cause RSSI to fluctuate. Suitable sampling rate can reduce foregoing problems.The noise that too small sampling rate reads proposed method to RFID It is more sensitive, and too big sampling rate makes posture obscurity boundary (that is, the sample of activity conversion at large may receive) between class.In In implementation, continuous RSSI data flow is acquired with 0.5 second sampling rate of ≈.

Data acquisition.For collection activity data, devise above by reference to described in Fig. 5 in people's daily routines most The posture and movement of 23 common orientation-sensitives.Six objects (1 women and 5 males) have participated in this experiment and each Object performs 23 particulates including 6 postures (above-mentioned posture (1)-(6)) and 17 movements (above-mentioned movement (7)-(23)) Spend movable set.Each object is required to execute each posture or movement 120 seconds.All 23 executed for object-order The different activity of class and the activity data acquired are considered as a set.So that data integration be it is publicly available, so as to again Show proposed scheme and provides support for other researchers in this field.

Authentication policy.Itd is proposed scheme is verified using (subject-dependent) model dependent on object, It is middle that the data acquired from each object are divided into the training dataset and be used to comment for being used for that feature selecting and identification model to be rebuild The test data set estimated.

For training dataset, such as execute the discribed segmentation step 530 of top half, the characteristic extraction step of Fig. 5 540, feature selection step 550, character subset generation step 560 and dictionary learning step 570, to be directed to every class activity structure Build an associated active dictionary.

For test data set, such as execute the discribed segmentation step 530 ' in lower half portion, the characteristic extraction step of Fig. 5 540 ', feature selection step 550 '.Also, the active dictionary created is assessed using test data set, to generate Assessment result.

The assessment result finally presented, i.e. accuracy rate (precision), recall rate (recall) and Fl scoring, are all 6 The average value of a object.

Hereinafter, carry out influence of the characterising parameter to activity recognition in conjunction with Fig. 7a-7c.We are tested to study this hair Influence of three parameters to activity recognition in bright method, i.e., preceding k feature, dictionary size d and training data and test The ratio of data.

The influence of selected feature sizes k.How many most useful and most effective feature of preceding k character control is used for It is fed in sorting algorithm.Fig. 7 a shows activity recognition accuracy rate, recall rate and F1 when dictionary size d is fixed as 8 and comments Point.In figure 7 a, abscissa indicates the component number of the feature vector in feature vector set.For example, in feature vector set O ={ o₁,…,o_mIncluding 7 feature vectors (m=7) and in the case where using 12 RFID label tags, feature vector o_iIn point The number of amount is 12, and the component number of whole feature vectors amounts to 84.The result shows that in most cases, it is special with using Sign complete or collected works compare, and a subset of feature complete or collected works is selected to improve classification performance.It could be observed that working as sorting algorithm most preferably When execution, with the increase (i.e. with the increase of component) of the number of selected feature, performance also increases, until the number of component Mesh reaches 34.

The influence of dictionary size d.Active dictionary is the excessively complete set of basis vector, and the number of vector indicates dictionary Size.Similar to research k influence experiment, in the case where fixed character size k=5 by dictionary size d from 2 change to 16, it is as a result shown in fig.7b.As can be seen that classification performance reaches highest at d=8 from Fig. 7 b, property retention later Stablize and is even slightly reduced as d=14.In view of both efficiency and accuracy, d=8 is set as default in experiment Dictionary size.

The influence of the ratio of training data and test data.The important factor of third for influencing activity recognition performance is that have How many training datas should participate in proposed model.By by entire data in the case where fixed k=5 and d=8 The training data ratio of collection, which changes from 0.05 to 0.7, to be assessed, and result is shown in figure 7 c.It is observed that only Using 10% sample for training, the method proposed reaches more than 65% accuracy rate, and it is only 20% Data as training data in the case where reach more than 90% accuracy rate.Property retention improve and when 40% data by with Performance reaches peak value when learning model.In an experiment, setting 0.2 is used as default trained percentage.

In short, an important advantage of the method proposed is that it only needs the training sample of very peanut can be with Desired horizontal executed activity recognition.Due to the difficulty in tag data, this right and wrong for mankind's activity identification Often with attractive feature.

Hereinafter, it will focus on following two aspect to describe the ratio for proposed method and other existing schemes Compared with experiment: i) proposed based on the method for sparse dictionary compared with four kinds based on the method for rarefaction representation and ii) mentioned Method out and other four kinds general classification being widely used in general fit calculation field (ubiquitous community) The performance of device compares.

Firstly, implementing four based on the rarefaction representation kind discovery learning scheme proposed in following two documents: (1) " the Robust face recognition via of J.Wright, A.Y.Yang, A.Ganesh, S.S.Sastry and Y.Ma sparse representation"；Pattern Analysis and Machine Intelligence,IEEE Transactions, 31 (2): page 210-227,2009；(2)J.J.Liu,W.Xu,M.-C.Huang,N.Alshurafa, " the A dense pressure sensitive bedsheet of M.Sarrafzadeh, N.Raut and B.Yadegar Design for unobtrusive sleep posture monitoring ", Pervasive Computing and Communications (PerCom), 2013IEEE International Conference, page 207-215, IEEE, 2013。

In above-mentioned discovery learning scheme, training sample is used directly to building dictionary.Assuming that by being converted from K activity Matrix B=[B that obtained training feature vector is formed₁... B_K], wherein B_iFor the subset of the training sample from movable i.It gives Determine query sample o^*, overall process includes following two key step.

First step is about finding out o^*Rarefaction representation on B.We calculate o^*Rarefaction representation, that is, from different activities Associated sparse coefficient vector w_i, o can be used for reconstruction^*.In other words, query sample o^*Can be represented as linearly across Away from

Above equation can be indicated again are as follows:

Wherein can be used the clean cut system newton interior point method recorded in the following documents, viaIt minimizes to find Sparse solution: " the An interior-point of S.-J.Kim, K.Koh, M.Lustig, S.Boyd and D.Gorinevsky method for large-scale l1-regularized least squares"；Selected Topics in Signal Processing, IEEE Journal, 1 (4): page 606-617,2007.Disclosure of the documents pass through reference Mode is fully incorporated in this.

Second step is about classification.Given query sample o^*, four kinds of heuristic learning methods using reconstructed coefficients, pass through Usage factorSubspace structure execute activity classification, be described in detail as follows.

Greatest coefficient (MC).Test sample label and o_iGreatest coefficient it is associated.Therefore, for the prediction of query sample Active tags l_iAre as follows:

Wherein, δ_k(o^*) indicate o^*In only with label l_kAssociated coefficient.

Greatest coefficient summation (MCS).The prediction label for searching sample is o^*Coefficient the label that is maximized of summation:

For each label l_k, calculate the o for corresponding to the training sample for belonging to each label^*Coefficient summation.For giving Fixed test RSSI, having the label of maximum total value is the activity being predicted.

Least residual (MR).For every class activity k, defined feature functionIt is selected and the The k associated coefficient of posture class.For It is new vector, only non-zero entry isIn with k-th The associated entry of posture class.It can be by given query sample o^*It is redeveloped intoIt therefore, can be with base In the reconstruction approximation from each activity class and by o^*It is categorized into true o^*With estimationBetween with least residual Posture class:

It is then possible to using following equation by o_iIt is classified as the movable k with Minimum Residual difference:

The maximum number of nonzero coefficient (referred to as " non-zero ").For inquiring RSSI sample o^*, the active tags quilt of prediction It indicates are as follows:

Wherein δ_k(o^*) indicate o^*In only with label l_kAssociated coefficient.| | indicate δ_k(o^*) length.

Second, the quality of the character subset of generation is assessed using multiple classifiers.Following sorting technique is selected, because it Activity recognition application is successfully applied in nearest document.

HaveMultinomial logistic regression (MLGL1) be amendment to linear regression, can logic-based function come Predict relied on variable.Multinomial or multivariable calculating are solved and being decomposed into a series of binary variables.ItsJust Then change using compensation (penalty) item, which makes the summation of the absolute value of parameter smaller, this typically results in Sparse parameter Vector.Herein, willRegularization is integrated into the linear classifier in target item.Given multiclass activity recognition problem, willRegularization is combined with multinomial logistic regression, to conditional probabilityIt is modeled.Just The main problem then changed can be calculated by optimizing log likelihood:

Nearest neighbor method (kNN) is the general category device for various classification problems.It trains example by K arest neighbors The majority voting of class label carrys out the class of forecast sample.

Linear SVM (LSVM), which is directed to, maximizes the super of the allowance between inhomogeneous supporting vector by determining Plane finds the optimal separation of binary mark example.The sequence of given training RSSI and corresponding posture labelWhereinLabel l ∈ 1 ..., k }.Objective function It is represented as:

s.t.l_i(w^Tφ(o_i)+b)≥1-ξ_i, i=1,2 ..., n

ξ_i>=0, i=1,2 ..., n (17) wherein ξ_iIt is slack variable, C is the compensation of error term, K (o_i, o_j)=φ (o_i)^Tφ(o_j) it is kernel function.

Random forest (RF) establishes the forest of decision tree, with same distribution but is independent from each other output class.At random Forest is based on the randomly choosing and based on the combination for the output set individually to feature for each tree.

Fig. 8 shows overall performance comparison result.As can be seen that the activity monitoring of the invention based on RFID is (referred to as " RFM ") every other method of obviously winning, this shows good potentiality and validity in activity classification.

Fig. 9 is shown in a manner of confusion matrix comments the detailed performance of the activity recognition method of embodiment according to the present invention Estimate.As can be seen that activity that most of mistake occurs to have similar class internal clearance in identification (for example, above-mentioned activity (22) to Left tumble and (23) are fallen to the right) when.As can be seen from the results, method of the invention can accurately identify mixed and disorderly indoor ring The activity of most of orientation-sensitive in border.

Hereinafter, by conjunction with Figure 10 a-10d come the feature selection approach based on CCA used by describing in the present invention With between the feature selection approach that is widely adopted at present, efficiency (for example, runing time) and validity (for example, accuracy rate/ Recall rate/Fl) in terms of comparative experiments.Specifically, it compares the forward direction selection method proposed based on CCA and is commented based on F The method of point (Fisher Score, FScore), based on having before the sequence for exempting F scoring to (Sequential Forward With Relief-F Score, SFRF) method and based on F statistics scoring forward direction select (Forward Selection with F-Statistics Score, SFSS) method.

Feature selecting based on FScore.Feature selecting based on FScore is for quantifying ith feature o_iScoring:

Wherein n_kIt is the number of the sample in k-th of activity class,And v_ikIt is the mean value and variance of ith feature respectively, AndIt is the mean value of ith feature.

Feature selecting based on SFRF.Based on the feature selection approach of SFRF according to the relevance values of feature in much journeys The data point of same campaign class and different activity classes close to each other is distinguished on degree to estimate the correlation of feature.This method calculates The weight of each feature is to quantify its value.According to following evaluation function for the sample of signal presented in each activity class To update the weight:

Wherein nearmiss_j(o_i) and nearhit_i(o_i) indicate respectively from the o from same campaign class and different activity classes_i Nearest RSSI sample.

Feature selecting based on SFSS.This method measures the difference of multiple real number sets, and following formula can be used It calculates:

Wherein n_jIt is the number of the sample in j-th of activity class,Indicate the mean value for the label i that training data is concentrated, and AndIt is the mean value of i-th of label in j-th of activity class.Difference between the positive set of molecule instruction and negative set, and point One in each set in two set of mother's instruction.F scoring is bigger, and this feature is more possible in the area activity recognition Zhong Shiyou Component.

The runing time of comparative feature selection course first, because identification delay is the critical consideration of activity recognition application. According to the essence of algorithm, runing time is divided into feature ordering time and characteristic set assessment time.It is of the invention based on The feature selecting of sequence, which executes entire characteristic set, initially to sort and therefore has constant runing time, the constant fortune Actual quantity of the row time independent of feature to be selected.Figure 10 a depicts four kinds based on the method for feature selecting entire Runing time on data acquisition system.It can be seen that and of the invention be based on other three kinds based on the feature selecting of CCA from Figure 10 a The method of feature selecting, which is compared, has competitive performance, even if the method based on FScore uses the least time.

In addition to efficiency, Figure 10 b- Figure 10 d presents the comparing result using four kinds of feature selection approach, and this hair The bright feature selection approach based on CCA is obviously won other three kinds of methods.Feature selection approach based on CCA of the invention is not The natural distribution of characteristic component is only remained, and discloses the mutual independence between characteristic component.In view of at runtime Between and accuracy in terms of the assessment result that combines, feature selection approach of the invention is with good performance in activity recognition.

In second aspect, embodiments of the present invention additionally provide a kind of activity recognition device.Figure 11 is shown according to this Invent the block diagram of the activity recognition device 1100 an of embodiment.As shown, device 1100 includes: signal acquisition unit 1010, it is configured as acquiring the movable radio frequency discrimination RFID signal of every class in the predefined activity of multiclass；Feature mentions Unit 1120 is taken, is configured as extracting multiple active characteristics from the RFID signal for every class activity；And study Unit 1130 is configured as learning the multiple active characteristics, is associated with constructing one for every class activity Active dictionary so that activity classification to be identified is one of predefined activity of the multiclass by constructed active dictionary.

In one embodiment, each active dictionary in constructed active dictionary includes for characterizing respective activity , the set of the basis vector of linear independence.

In one embodiment, device 1100 further comprises: correlation calculations unit, is configured as described every Class activity calculates the correlation between every a pair of of active characteristics in the multiple active characteristics；And feature selection unit, quilt It is configured to correlation calculated and selects the active characteristics of predetermined number from the multiple active characteristics.Unit 1130 are further configured to learn selected active characteristics, to construct movable word associated with respective activity Allusion quotation.

In one embodiment, feature selection unit is further configured to: according to correlation calculated from as low as High sequence is ranked up the multiple active characteristics；And it is selected from ranked multiple active characteristics described predetermined The active characteristics of number.

In one embodiment, device 1100 further comprises taxon, which is configured as: in response to It receives from the movable inquiry RFID signal to be identified, it is special that multiple activities is extracted from the inquiry RFID signal Sign, to generate query feature vector set；Using the set of the basis vector in each active dictionary come respectively to described Query feature vector set carries out sparse coding, to generate the signal of multiple reconstructions；By the query feature vector set and institute Each of the signal for stating multiple reconstructions is compared；And in response to the query feature vector set with it is the multiple heavy One of signal built matches, and is activity associated with the signal of matched reconstruction by the activity classification to be identified.

In one embodiment, which is further configured to: calculate the query feature vector set with Reconstruction error between each of signal of the multiple reconstruction；And in response to the query feature vector set and institute The reconstruction error stated between one of signal of multiple reconstructions is minimum, is to miss with the smallest reconstruction by the activity classification to be identified The associated activity of difference.

In one embodiment, feature extraction unit 1120 is further configured to extract multiple statistical natures.

In one embodiment, multiple statistical natures include at least three in the following terms: the radio frequency identification The minimum value of RFID signal；The maximum value of the radio frequency discrimination RFID signal；The mean value of the radio frequency discrimination RFID signal；It is described The variance of radio frequency discrimination RFID signal；The root mean square of the radio frequency discrimination RFID signal；The mark of the radio frequency discrimination RFID signal It is quasi- poor；And the intermediate value of the radio frequency discrimination RFID signal.

In one embodiment, signal acquisition unit 1110 be further configured to from RFID receiver acquire from The movable RFID signal of every class, and execute and do not deposited between the movable object of every class and the RFID receiver It is being physically contacted.

Particularly, embodiment according to the present invention may be implemented as calculating above with reference to Fig. 2-10 process described Machine software program.For example, embodiments of the present invention include a kind of computer program product comprising be tangibly embodied in machine Computer program on readable medium, the computer program include the program code for executing method 400.

In general, various example embodiments of the invention can hardware or special circuit, software, logic or its Implement in any combination.Some aspects can be implemented within hardware, and other aspects can be can be by controller, microprocessor Or other calculate and implement in the firmware or software that equipment executes.When the various aspects of embodiments of the present invention are illustrated or described as When block diagram, flow chart or other certain graphical representations of use, it will be understood that box described herein, device, system, techniques or methods Can be used as unrestricted example hardware, software, firmware, special circuit or logic, common hardware or controller or other It calculates and implements in equipment or its certain combination.

Moreover, each frame in flow chart can be counted as method and step and/or the operation of computer program code generates Operation, and/or be interpreted as execute correlation function multiple couplings logic circuit component.For example, embodiments of the present invention Including computer program product, which includes the computer journey visibly realized on a machine-readable medium Sequence, the computer program include the program code for being configured as realizing above description method.

In disclosed context, machine readable media can be include or storage be used for or about instruction execution system Any tangible medium of the program of system, device or equipment.Machine readable media can be machine-readable signal medium or machine can Read storage medium.Machine readable media can include but is not limited to electronics, magnetic, it is optical, electromagnetism, infrared or partly lead Body system, device or equipment equipment or its any appropriate combination.The more detailed example of machine readable storage medium includes having one Or the electrical connection of multiple conducting wires, portable computer diskette, hard disk, random access memories (RAM), read-only memory (ROM), Erasable Programmable Read Only Memory EPROM (EPROM or flash memory), light storage device, magnetic storage apparatus or its is any appropriate Combination.

It can be write with one or more programming languages for realizing the computer program code of method of the invention.These Computer program code can be supplied to the processing of general purpose computer, special purpose computer or other programmable data processing units Device so that program code when being executed by computer or other programmable data processing units, cause flow chart and/ Or function/operation specified in block diagram is carried out.Program code can completely on computers, part on computers, conduct Independent software package, part are on computers and part is held on a remote computer or server on the remote computer or completely Row.

In addition, although operation is depicted with particular order, this simultaneously should not be construed and require this generic operation to show Particular order is completed with sequential order, or executes the operation of all diagrams to obtain expected result.In some cases, more Task or parallel processing can be beneficial.Similarly, although discussed above contain certain specific implementation details, this is not It should be interpreted that any invention of limitation or the scope of the claims, and should be interpreted that the particular implementation side that can be directed to specific invention The description of formula.Certain features described in the context of separated embodiment can also be with combined implementation in list in this specification In a embodiment.On the contrary, the various features described in the context of single embodiment can also be discretely multiple Embodiment is implemented in any suitable sub-portfolio.

For aforementioned example embodiment of the invention various modifications, change will be when checking foregoing description together with attached drawing Obvious are become to those skilled in the technology concerned.Any and all modification will still fall within unrestricted and of the invention example Embodiment range.In addition, aforementioned specification and attached drawing have the benefit inspired, it is related to the skill of these embodiments of the invention The technical staff in art field will be appreciated that the other embodiments of the invention illustrated herein.

It will be understood that embodiments of the present invention are not limited to disclosed particular implementation, and modify and other implementations Mode should be all contained in scope of the appended claims.Although being used here specific term, they are only general It is used in the sense that description, and is not limited to purpose.

Claims (16)

1. a kind of activity recognition method, comprising:

Acquire the movable radio frequency discrimination RFID signal of every class in the predefined activity of multiclass；

For every class activity, multiple active characteristics are extracted from the RFID signal；And

The multiple active characteristics are learnt, to construct an associated active dictionary for every class activity, are made It is one of predefined activity of the multiclass that constructed active dictionary, which is obtained, by activity classification to be identified；

It is wherein that one of predefined activity of the multiclass includes: by activity classification to be identified

In response to receiving from the movable inquiry RFID signal to be identified, extracted from the inquiry RFID signal Multiple active characteristics, to generate query feature vector set；

Using the set of the basis vector in each active dictionary in constructed active dictionary, come special to the inquiry respectively It levies vector set and carries out sparse coding, to generate the signal of multiple reconstructions；

Each of the query feature vector set and the signal of the multiple reconstruction are compared；And

Match in response to one of the query feature vector set and the signal of the multiple reconstruction, by the work to be identified It is dynamic be classified as with the associated activity of the signal of matched reconstruction.

2. according to the method described in claim 1, wherein each active dictionary in constructed active dictionary includes being used for table The set of basis vector that levy respective activity, linear independence.

3. according to the method described in claim 2, further comprising:

For every class activity, the correlation between every a pair of of active characteristics in the multiple active characteristics is calculated；

The active characteristics of predetermined number are selected from the multiple active characteristics based on the calculated correlation；And

Selected active characteristics are learnt, to construct active dictionary associated with respective activity.

4. according to the method described in claim 3, wherein being selected from the multiple active characteristics based on the calculated correlation The active characteristics for selecting predetermined number include:

According to the sequence of correlation calculated from low to high, the multiple active characteristics are ranked up；And

The active characteristics of the predetermined number are selected from ranked multiple active characteristics.

5. according to the method described in claim 1, wherein by the signal of the query feature vector set and the multiple reconstruction Each of to be compared include: each of the signal for calculating the query feature vector set and the multiple reconstruction Between reconstruction error；And

It include: wherein in response to institute for activity associated with the signal of matched reconstruction by the activity classification to be identified The reconstruction error stated between one of signal of query feature vector set and the multiple reconstruction is minimum, by the work to be identified It is dynamic to be classified as activity associated with the smallest reconstruction error.

6. method according to claim 1 to 4, wherein extracting multiple active characteristics includes extracting multiple statistics Feature.

7. according to the method described in claim 6, wherein the multiple statistical nature includes at least three in the following terms:

The minimum value of the radio frequency discrimination RFID signal；

The maximum value of the radio frequency discrimination RFID signal；

The mean value of the radio frequency discrimination RFID signal；

The variance of the radio frequency discrimination RFID signal；

The root mean square of the radio frequency discrimination RFID signal；

The standard deviation of the radio frequency discrimination RFID signal；And

The intermediate value of the radio frequency discrimination RFID signal.

8. method according to claim 1 to 4, wherein living from every class from RFID receiver acquisition The dynamic RFID signal；And

It wherein executes and physical contact is not present between the movable object of every class and the RFID receiver.

9. a kind of activity recognition device, comprising:

Signal acquisition unit is configured as acquiring the movable radio frequency discrimination RFID of every class in the predefined activity of multiclass Signal；

Feature extraction unit is configured as extracting multiple active characteristics from the RFID signal for every class activity；

Unit is configured as learning the multiple active characteristics, to construct a phase for every class activity Associated active dictionary, so that activity classification to be identified is the predefined activity of the multiclass by constructed active dictionary One；And

Taxon is configured as:

In response to receiving from the movable inquiry RFID signal to be identified, extracted from the inquiry RFID signal Multiple active characteristics, to generate query feature vector set；

Using the set of the basis vector in each active dictionary in constructed active dictionary, come special to the inquiry respectively It levies vector set and carries out sparse coding, to generate the signal of multiple reconstructions；

Each of the query feature vector set and the signal of the multiple reconstruction are compared；And

Match in response to one of the query feature vector set and the signal of the multiple reconstruction, by the work to be identified It is dynamic be classified as with the associated activity of the signal of matched reconstruction.

10. device according to claim 9, wherein each active dictionary in constructed active dictionary includes being used for table The set of basis vector that levy respective activity, linear independence.

11. device according to claim 10, further comprises:

Correlation calculations unit is configured as calculating every a pair of living in the multiple active characteristics for every class activity Correlation between dynamic feature；And

Feature selection unit is configured as selecting predetermined number from the multiple active characteristics based on the calculated correlation Active characteristics；

Wherein the unit is further configured to learn selected active characteristics, with building and respective activity Associated active dictionary.

12. device according to claim 11, wherein the feature selection unit is further configured to:

According to the sequence of correlation calculated from low to high, the multiple active characteristics are ranked up；And

The active characteristics of the predetermined number are selected from ranked multiple active characteristics.

13. device according to claim 9, wherein the taxon is further configured to:

Calculate the reconstruction error between each of signal of the query feature vector set and the multiple reconstruction；And

It is minimum in response to the reconstruction error between one of the query feature vector set and the signal of the multiple reconstruction, by institute Stating activity classification to be identified is activity associated with the smallest reconstruction error.

14. the device according to any one of claim 9 to 12, wherein the feature extraction unit is further configured to Extract multiple statistical natures.

15. device according to claim 14, wherein the multiple statistical nature includes at least three in the following terms:

The minimum value of the radio frequency discrimination RFID signal；

The maximum value of the radio frequency discrimination RFID signal；

The mean value of the radio frequency discrimination RFID signal；

The variance of the radio frequency discrimination RFID signal；

The root mean square of the radio frequency discrimination RFID signal；

The standard deviation of the radio frequency discrimination RFID signal；And

The intermediate value of the radio frequency discrimination RFID signal.

16. the device according to any one of claim 9 to 12, wherein the signal acquisition unit is further configured to From RFID receiver acquire from every movable RFID signal of class, and execute the movable object of every class with There is no physical contacts between the RFID receiver.