CN105046193B - A kind of human motion recognition method based on fusion rarefaction representation matrix - Google Patents

Abstract

The invention discloses a kind of human motion recognition methods based on fusion rarefaction representation matrix, including gathered data, the action matrix step of each test object, the rarefaction representation matrix step for building complete dictionary step, solving each test object are built, the first fusion rarefaction representation matrix step is solved, solves the second fusion rarefaction representation matrix and action recognition step.The present invention is trained using a large amount of different human action datas, pass through fusion methods different twice, the error of rarefaction representation matrix that same target in training process obtains is reduced first, then reduce the difference between different objects, reduce redundancy, effective information during integrated classification is optimized complementary information and cooperative information, improves the accuracy rate of human action identification.

Description

A kind of human motion recognition method based on fusion rarefaction representation matrix

Technical field

The present invention relates to a kind of human motion recognition method, more particularly to a kind of human body based on fusion rarefaction representation matrix Action identification method.

Background technology

Wireless body area network (Wireless Body Area Network, WBAN) is a kind of short distance centered on human body Cordless communication network.By wearing the different sensor of function on human body, WBAN, which becomes personal physiological information, to be acquired and transmits One of important technical, obtained in fermentation such as daily life monitoring, patient care, sportsman's rehabilitation supplemental trainings Extensive use, wherein, the human action identification based on the inertial sensor worn in WBAN becomes current research hotspot.At present Common human motion recognition method mainly has decision tree, artificial neural network, support vector machines, multilevel hierarchy method, sparse table Show classification etc..At present, the technology research is merged to information very big breakthrough, but at present still without one A definition merged about information being widely recognized as, according to a variety of different definition, it is real that we can be understood as information fusion Matter is exactly to make full use of multiple sensors or the perception data of multi-characteristicattribute attribute, is reasonably used with certain rule and dominates this A little perception datas, comprehensive multi-faceted incomplete local environmental information, eliminate contradiction that may be present between information, obtain to sense Know the description or explanation of object.As it can be seen that information fusion process is substantially the process of an information processing, the mode of processing has Pixel-based fusion, feature-based fusion and decision level fusion.The core of information fusion is to the coordination optimization of sensing data and comprehensive Conjunction handle, by information fusion technology be applied to human action identification in, can efficiently reduce multiple inertial sensor datas it Between redundancy, optimize cooperative information and complementary information, be conducive to the promotion of human action discrimination.

Invention content

The purpose of the present invention is to provide a kind of human motion recognition methods based on fusion rarefaction representation matrix.

To achieve these goals, the present invention adopts the following technical scheme that：

A kind of human motion recognition method based on fusion rarefaction representation matrix, includes the following steps：

Step 1：K test object is selected, each test object wears L sensor respectively, and each sensor is adopted The signal kinds of collection are identical, and including acceleration signal and gyroscope signal, each test object does different action types Number is T, and each test object is N to the number of repetition of the various actions, and each sensor is to various actions Repeat for each time carry out h time and sample, generate the action matrix H of each test object_p∈R^U×W, p=1,2 ..., K：

H_p=[H_p,1,H_p,2,...,H_p,T]∈R^U×W (1)

Wherein, the action submatrix H of U=5hL, W=NT, p-th people q kinds action_p,qFor：

H_p,q=[v_p,q,1,v_p,q,2,...,v_p,q,N]∈R^U×N (2)

The action vector v of wherein p-th people n-th q kinds action_p,q,nFor：

Wherein, p-th of people n-th q kind acts the corresponding action subvector β of j-th of sensor_p,q,n,j, j=1, 2 ..., L is：

β_p,q,n,j=(a_p,q,n,j(1)^T,a_p,q,n,j(2)^T,...,a_p,q,n,j(h)^T)^T∈R^5h (4)

Wherein, p-th of people n-th q kind acts the sensor vector a that j-th of sensor is acquired in t moment_p,q,n,j(t) For：

a_p,q,n,j(t)=(x_p,q,n,j(t),y_p,q,n,j(t),z_p,q,n,j(t),θ_p,q,n,j(t),ρ_p,q,n,j(t))^T∈R⁵ (5)

Wherein x_p,q,n,j(t),y_p,q,n,j(t),z_p,q,n,j(t) j-th of sensing of q kinds action is done for p-th of people's n-th Device is in t moment collected X, Y and the acceleration signal of Z-direction, θ_p,q,n,j(t),ρ_p,q,n,j(t) it is q for p-th of people's n-th J-th of sensor of kind action is in the collected two axis gyroscope instrument X of t moment, the angular velocity signal of Y-direction；

Step 2：The action submatrix H of the various actions is by each test object_p,qGenerated complete dictionary matrix A：

A=[H_1,1,H_2,1,...,H_K,1,H_1,2,H_2,2,...,H_K,2,...,H_1,T,H_2,T,...,H_K,T]∈R^U×Q (6)

Wherein, Q=NTK；

Step 3：According to the excessively complete dictionary matrix A, using minimization L1 norms under quadratic constraints method generation with The action matrix H of each test object_p∈R^U×W, the rarefaction representation matrix B of the corresponding each test object of p=1,2 ..., K^p ∈R^Q×W, p=1,2 ..., K；

Step 4：By the rarefaction representation matrix B of each test object^p∈R^Q×W, p=1,2 ..., K are maximized by cycle L1 norm methods extract the first fusion rarefaction representation Matrix C of each test object^p∈R^Q×T, p=1,2 ..., K；Described first melts Close rarefaction representation Matrix C^p∈R^Q×TColumn vector number and it is described action type number T-phase it is same；

Step 5：Rarefaction representation Matrix C is merged by the first of each test object^p∈R^Q×T, p=1,2 ..., K pass through Cycle maximizes L1 norm methods and obtains the second fusion rarefaction representation matrix F ∈ R^Q×T, the second fusion rarefaction representation matrix F∈R^Q×TColumn vector be f_q, q=1,2 ..., T are corresponding with the various actions respectively；

Step 6：Carry out human action identification：Calculate action vector γ to be identified_testRarefaction representation is merged with described second The linear weighted function difference of each column vector in matrix F, select action type corresponding to the column vector of linear weighted difference minimum for Final recognition result.

Each test object wears 5 sensors respectively, and the signal kinds of each sensor acquisition are identical, include Acceleration signal and gyroscope signal, each test object do 13 kinds of different actions, each action is done 5 times.

The method with minimization L1 norms under quadratic constraints described in the step 3 comprises the steps of：

Step 3-1：Solve the action matrix H of p-th of test object_p∈R^U×W, p=1,2 ..., each column vector of K Corresponding rarefaction representation vectorN=1,2 ..., N.：

Wherein ε is observation noise.

Step 3-2：By the action matrix H of p-th of test object_p∈R^U×W, p=1,2 ..., each column vector pair of K The rarefaction representation matrix B that the rarefaction representation vector answered is formed^p, p=1,2 ..., K：

The step 4 is made of step in detail below：

Step 4-1：By the rarefaction representation matrix B of each test object^p, p=1,2 ..., K column major orders are equally divided into With the action type number T-phase with first to T sparse vector groups, every group of N number of vector：

Step 4-2：Seek the optimal column vector of each sparse vector group：It enables Q=1,2 ..., T, n=1,2 ..., N；

Step 4-3：By the optimal column vector of each sparse vector groupQ=1,2 ..., T, n=1, 2 ..., N forms the sparse output matrix C of the second fusion in order^p=[d^p,1,d^p,2,...,d^p,T]∈R^Q×T, p=1,2 ..., K.

The step 5 is made of step in detail below：

Step 5-1：Rarefaction representation Matrix C is merged by the first of each test object^p∈R^Q×T, p=1,2 ..., K are pressed Row sequence is equally divided into the T+1 same with the action type number T-phase to 2T sparse vector groups, every group of K vector：

(d^1,1,d^2,1,...,d^K,1), (d^1,2,d^2,2,...,d^K,2) ..., (d^1,T,d^2,T,...,d^K,T)；

Step 5-3：The optimal column vector of T+1 to the 2T sparse vector groups is formed into second fusion in order Rarefaction representation matrix F=[f₁,f₂,...,f_T]∈R^Q×T。

The difference r of linear weighted function described in the step 6_q(γ_test) computational methods be：

r_q(γ_test)=| | γ_test-f_q||₂, q=1,2 ..., T (9)

The beneficial effects of the present invention are：

The present invention is trained using a large amount of different human action datas, by fusion methods different twice, first The error of rarefaction representation matrix that same target in training process obtains is reduced, then reduces the difference between different objects, Reduce redundancy, the effective information during integrated classification is optimized complementary information and cooperative information, improves people The accuracy rate of body action recognition.

Description of the drawings

Fig. 1 is the flow chart of the present invention.

Specific embodiment

Technical solution for a better understanding of the present invention, below will be to of the invention a kind of dilute based on merging with reference to attached drawing The human motion recognition method for dredging representing matrix is described in further detail, and system block diagram is as shown in Figure 1, specific implementation step It is as follows：

A kind of human motion recognition method based on fusion rarefaction representation matrix, includes the following steps：

Step 1：K test object is selected, each test object wears L sensor respectively, and each sensor is adopted The signal kinds of collection are identical, and including acceleration signal and gyroscope signal, each test object does different action types Number is T, and each test object is N to the number of repetition of the various actions, and each sensor is to various actions Repeat for each time carry out h time and sample, generate the action matrix H of each test object_p∈R^U×W, p=1,2 ..., K：

H_p=[H_p,1,H_p,2,...,H_p,T]∈R^U×W (1)

Wherein, the action submatrix H of U=5hL, W=NT, p-th people q kinds action_p,qFor：

H_p,q=[v_p,q,1,v_p,q,2,...,v_p,q,N]∈R^U×N (2)

The action vector v of wherein p-th people n-th q kinds action_p,q,nFor：

Wherein, p-th of people n-th q kind acts the corresponding action subvector β of j-th of sensor_p,q,n,j, j=1, 2 ..., L is：

β_p,q,n,j=(a_p,q,n,j(1)^T,a_p,q,n,j(2)^T,...,a_p,q,n,j(h)^T)^T∈R^5h (4)

Wherein, p-th of people n-th q kind acts the sensor vector a that j-th of sensor is acquired in t moment_p,q,n,j(t) For：

a_p,q,n,j(t)=(x_p,q,n,j(t),y_p,q,n,j(t),z_p,q,n,j(t),θ_p,q,n,j(t),ρ_p,q,n,j(t))^T∈R⁵ (5)

Wherein x_p,q,n,j(t),y_p,q,n,j(t),z_p,q,n,j(t) j-th of sensing of q kinds action is done for p-th of people's n-th Device is in t moment collected X, Y and the acceleration signal of Z-direction, θ_p,q,n,j(t),ρ_p,q,n,j(t) it is q for p-th of people's n-th J-th of sensor of kind action is in the collected two axis gyroscope instrument X of t moment, the angular velocity signal of Y-direction；

Step 2：The action submatrix H of the various actions is by each test object_p,qGenerated complete dictionary matrix A：

A=[H_1,1,H_2,1,...,H_K,1,H_1,2,H_2,2,...,H_K,2,...,H_1,T,H_2,T,...,H_K,T]∈R^U×Q (6)

Wherein, Q=NTK；

Step 3：According to the excessively complete dictionary matrix A, using minimization L1 norms under quadratic constraints method generation with The action matrix H of each test object_p∈R^U×W, the rarefaction representation matrix B of the corresponding each test object of p=1,2 ..., K^p ∈R^Q×W, p=1,2 ..., K；

Step 4：By the rarefaction representation matrix B of each test object^p∈R^Q×W, p=1,2 ..., K are maximized by cycle L1 norm methods extract the first fusion rarefaction representation Matrix C of each test object^p∈R^Q×T, p=1,2 ..., K；Described first melts Close rarefaction representation Matrix C^p∈R^Q×TColumn vector number and it is described action type number T-phase it is same；

Step 5：Rarefaction representation Matrix C is merged by the first of each test object^p∈R^Q×T, p=1,2 ..., K pass through Cycle maximizes L1 norm methods and obtains the second fusion rarefaction representation matrix F ∈ R^Q×T, the second fusion rarefaction representation matrix F∈R^Q×TColumn vector be f_q, q=1,2 ..., T are corresponding with the various actions respectively；

Step 6：Carry out human action identification：Calculate action vector γ to be identified_testRarefaction representation is merged with described second The linear weighted function difference of each column vector in matrix F, select action type corresponding to the column vector of linear weighted difference minimum for Final recognition result.

Each test object wears 5 sensors respectively, and the signal kinds of each sensor acquisition are identical, include Acceleration signal and gyroscope signal, each test object do 13 kinds of different actions, each action is done 5 times.

The method with minimization L1 norms under quadratic constraints described in the step 3 comprises the steps of：

Step 3-1：Solve the action matrix H of p-th of test object_p∈R^U×W, p=1,2 ..., each column vector of K Corresponding rarefaction representation vectorN=1,2 ..., N.：

Wherein ε is observation noise, takes 0.01 in the present embodiment.

Step 3-2：By the action matrix H of p-th of test object_p∈R^U×W, p=1,2 ..., each column vector pair of K The rarefaction representation matrix B that the rarefaction representation vector answered is formed^p, p=1,2 ..., K：

The step 4 is made of step in detail below：

Step 4-1：By the rarefaction representation matrix B of each test object^p, p=1,2 ..., K column major orders are equally divided into With the action type number T-phase with first to T sparse vector groups, every group of N number of vector：

Step 4-2：Seek the optimal column vector of each sparse vector group：It enables Q=1,2 ..., T, n=1,2 ..., N；

Step 4-3：By the optimal column vector of each sparse vector groupQ=1,2 ..., T, n=1, 2 ..., N forms the sparse output matrix C of the second fusion in order^p=[d^p,1,d^p,2,...,d^p,T]∈R^Q×T, p=1,2 ..., K.

The step 5 is made of step in detail below：

Step 5-1：Rarefaction representation Matrix C is merged by the first of each test object^p∈R^Q×T, p=1,2 ..., K are pressed Row sequence is equally divided into the T+1 same with the action type number T-phase to 2T sparse vector groups, every group of K vector：

(d^1,1,d^2,1,...,d^K,1), (d^1,2,d^2,2,...,d^K,2) ..., (d^1,T,d^2,T,...,d^K,T)；

Step 5-3：The optimal column vector of T+1 to the 2T sparse vector groups is formed into second fusion in order Rarefaction representation matrix F=[f₁,f₂,...,f_T]∈R^Q×T。

The difference r of linear weighted function described in the step 6_q(γ_test) computational methods be：

r_q(γ_test)=| | γ_test-f_q||₂, q=1,2 ..., T (9)

In conclusion the present invention provides a kind of human motion recognition method based on fusion rarefaction representation matrix, pass through Fusion methods different twice reduces the error of rarefaction representation matrix that same target in training process obtains, then first Reduce the difference between different objects, reduce redundancy, the effective information during integrated classification makes complementary information and association It is optimized with information, improves the accuracy rate of human action identification.

Claims (6)

1. a kind of human motion recognition method based on fusion rarefaction representation matrix, it is characterised in that：Include the following steps：

Step 1：K test object is selected, each test object wears L sensor respectively, each sensor acquisition Signal kinds are identical, and including acceleration signal and gyroscope signal, each test object does different action type numbers It is T, each test object is N to the number of repetition of the various actions, and each sensor is to each of various actions Secondary repetition carries out h sampling, generates the action matrix H of each test object_p∈R^U×W, p=1,2 ..., K：

H_p=[H_p,1,H_p,2,…,H_p,T]∈R^U×W (1)

Wherein, the action submatrix H of U=5hL, W=NT, p-th people q kinds action_p,qFor：

H_p,q=[v_p,q,1,v_p,q,2,…,v_p,q,N]∈R^U×N (2)

The action vector v of wherein p-th people n-th q kinds action_p,q,nFor：

Wherein, p-th of people n-th q kind acts the corresponding action subvector β of j-th of sensor_p,q,n,j, j=1,2 ..., L are：

Wherein, p-th of people n-th q kind acts the sensor vector a that j-th of sensor is acquired in t moment_p,q,n,j(t) it is：

a_p,q,n,j(t)=(x_p,q,n,j(t),y_p,q,n,j(t),z_p,q,n,j(t),θ_p,q,n,j(t),ρ_p,q,n,j(t))^T∈R⁵ (5)

Wherein x_p,q,n,j(t),y_p,q,n,j(t),z_p,q,n,j(t) j-th of sensor of q kinds action is made in t for p-th of people's n-th The acceleration signal of moment collected X, Y and Z-direction, θ_p,q,n,j(t),ρ_p,q,n,j(t) q kinds are done for p-th of people's n-th to move J-th of the sensor made is in the collected two axis gyroscope instrument X of t moment, the angular velocity signal of Y-direction；

Step 2：The action submatrix H of the various actions is by each test object_p,qGenerated complete dictionary matrix A：

A=[H_1,1,H_2,1,…,H_K,1,H_1,2,H_2,2,…,H_K,2,…,H_1,T,H_2,T,…,H_K,T]∈R^U×Q (6)

Wherein, Q=NTK；

Step 3：According to the excessively complete dictionary matrix A, using minimization L1 norms under quadratic constraints method generation with it is described The action matrix H of each test object_p∈R^U×W, the rarefaction representation matrix B of the corresponding each test object of p=1,2 ..., K^p∈R^Q×W, P=1,2 ..., K；

Step 4：By the rarefaction representation matrix B of each test object^p∈R^Q×W, p=1,2 ..., K maximize L1 models by cycle Number method extracts the first fusion rarefaction representation Matrix C of each test object^p∈R^Q×T, p=1,2 ..., K；First fusion is dilute Dredge representing matrix C^p∈R^Q×TColumn vector number and it is described action type number T-phase it is same；

Step 5：Rarefaction representation Matrix C is merged by the first of each test object^p∈R^Q×T, p=1,2 ..., K are by recycling most Bigization L1 norm methods obtain the second fusion rarefaction representation matrix F ∈ R^Q×T, the second fusion rarefaction representation matrix F ∈ R^Q×T's Column vector is f_q, q=1,2 ..., T are corresponding with the various actions respectively；

Step 6：Carry out human action identification：Calculate action vector γ to be identified_testRarefaction representation matrix F is merged with described second In each column vector linear weighted function difference, it is final to select action type corresponding to the column vector of linear weighted difference minimum Recognition result.

2. the human motion recognition method according to claim 1 based on fusion rarefaction representation matrix, it is characterised in that：

Each test object wears 5 sensors respectively, and the signal kinds of each sensor acquisition are identical, include accelerating Signal and gyroscope signal are spent, each test object does 13 kinds of different actions, each action is done 5 times.

3. the human motion recognition method according to claim 1 based on fusion rarefaction representation matrix, it is characterised in that：

The method with minimization L1 norms under quadratic constraints described in the step 3 comprises the steps of：

Step 3-1：Solve the action matrix H of p-th of test object_p∈R^U×W, p=1,2 ..., each column vector of K corresponds to Rarefaction representation vectorN=1,2 ..., N,

Wherein ε is observation noise；

Step 3-2：By the action matrix H of p-th of test object_p∈R^U×W, p=1,2 ..., each column vector of K is corresponding The rarefaction representation matrix B that rarefaction representation vector is formed^p, p=1,2 ..., K：

4. the human motion recognition method according to claim 1 based on fusion rarefaction representation matrix, it is characterised in that：

The step 4 is made of step in detail below：

Step 4-1：By the rarefaction representation matrix B of each test object^p, p=1,2 ..., K column major orders be equally divided into it is described Act type number T-phase with first to T sparse vector groups, every group of N number of vector：

Step 4-2：Seek the optimal column vector of each sparse vector group：It enables Q=1,2 ..., T, n=1,2 ..., N；

Step 4-3：By the optimal column vector of each sparse vector groupQ=1,2 ..., T, n=1,2 ..., N The sparse output matrix C of the second fusion of composition in order^p=[d^p,1,d^p,2,…,d^p,T]∈R^Q×T, p=1,2 ..., K.

5. the human motion recognition method according to claim 1 based on fusion rarefaction representation matrix, it is characterised in that：

The step 5 is made of step in detail below：

Step 5-1：Rarefaction representation Matrix C is merged by the first of each test object^p∈R^Q×T, p=1,2 ..., K column major orders The T+1 same with the action type number T-phase is equally divided into 2T sparse vector groups, every group of K vector：(d^1,1,d^{2 ,1},…,d^K,1), (d^1,2,d^2,2,…,d^K,2) ..., (d^1,T,d^2,T,…,d^K,T)；

Step 5-2：The T+1 is sought to the optimal column vector of 2T sparse vector groups：Enable A_q,p=| | d^q,p||₁,Wherein p=1,2 ..., K, q=1,2 ..., T；

Step 5-3：It is sparse that the optimal column vector of T+1 to the 2T sparse vector groups is formed into second fusion in order Representing matrix F=[f₁,f₂,…,f_T]∈R^Q×T。

6. the human motion recognition method according to claim 1 based on fusion rarefaction representation matrix, it is characterised in that：

The difference r of linear weighted function described in the step 6_q(γ_test) computational methods be：

r_q(γ_test)=| | γ_test-f_q||₂, q=1,2 ..., T (9).