阅读说明：本技术 一种基于微惯性的无器械健身动作质量评估方法 (Micro-inertia-based non-apparatus body-building action quality evaluation method ) 是由 阳媛 杨浩然 王庆 王慧青 况余进 于 2021-08-16 设计创作，主要内容包括：本发明公开了一种基于微惯性的无器械健身动作质量评估方法,可以对人的无器械健身运动进行识别与质量评估。该方法采用状态-动作两步分类方法,首先利用3个6轴微惯性传感器识别人体运动状态,然后调用Elman-Kalman轨迹估计模型,预测每个传感器节点的运动轨迹,再将轨迹信号与采集信号用于第二步健身动作分类。完成分类后,将生成的运动序列与标准运动序列库中的序列进行对比分析,评估动作的全身和局部的标准度与稳定度,并给出动作质量评估结果,用以帮助人们高效、安全地开展徒手健身活动。本发明在人的徒手运动健身过程中,对用户的健身动作类型能进行有效识别,准确率高,并能对用户身体各部分的运动姿态做出合理评估。(The invention discloses a micro-inertia based non-equipment fitness motion quality evaluation method, which can be used for identifying and evaluating the quality of non-equipment fitness motion of people. The method adopts a state-action two-step classification method, firstly, 3 6-axis micro-inertial sensors are utilized to identify the motion state of a human body, then an Elman-Kalman trajectory estimation model is called to predict the motion trajectory of each sensor node, and then a trajectory signal and an acquired signal are used for the second step of body-building action classification. After classification is completed, the generated motion sequence is compared with sequences in a standard motion sequence library for analysis, the whole body and local standard degree and stability of the motion are evaluated, and the motion quality evaluation result is given to help people to efficiently and safely carry out free-hand fitness activities. The method can effectively identify the body-building action type of the user in the bare-handed movement body-building process of the person, has high accuracy, and can reasonably evaluate the movement posture of each part of the body of the user.)

1. A method for evaluating the quality of non-equipment fitness actions based on micro-inertia is characterized by comprising the following steps:

(1) collecting and processing body-building exercise information;

in the step (1), the information of the body-building exercise is collected and processed, 6-axis inertial sensors are arranged on the human body, N inertial sensors are arranged on the human body, and the number of the inertial sensors is N, wherein the number of the upper body is N_uLower part of the body N_d1 in the waist and abdomen;

(2) a human body fitness state-action two-step classification model;

the two-step classification model of the fitness action without the apparatus comprises the following specific steps:

(2.1) the acceleration angle speed and the solved Euler angle information that utilize each sensor of upper part of the body, lower part of the body, a sensor of waist abdomen, totally 3 little inertial sensor collection of 6 axles carry out coarse classification to human motion state, divide into five types with the motion state, specifically do: upper limb movement, lower limb movement, waist and abdomen movement, whole body movement and non-movement state;

(2.2) calling the Elman-Kalman model in the step (3) according to the state classification result to obtain coordinate track information of each sensor, performing fine classification of body-building exercise according to the attitude angle information of all nodes, and specifically classifying the exercise state into various actions:

(2.3) for the sensor i, combining the classification result, the action track coordinate and the Euler angle to form an action quality evaluation sequence, wherein the specific form is as follows:

wherein x, y, z are three-dimensional coordinates of the attachment part of the motion process sensor, α, β, γ are corresponding Euler angle sequences, C₁The state classification result is evaluated as 1-5, which respectively correspond to 5 motion states, C₂Taking the value of 1-m as an action classification result, respectively corresponding to m actions in each motion state, and updating the value after the action classification step is finished;

(3) an Elman-Kalman trajectory estimation model of human motion;

the Elman-Kalman trajectory estimation model comprises the following specific steps:

(3.1) aiming at the first four motion states of the state classification result in the step (2), establishing 4 corresponding motion track estimation models for each sensor;

(3.2) the Elman-Kalman trajectory estimation model structure is specifically that in a hidden layer of an Elman neural network, estimation on future data is added, and an extended Kalman filter is constructed in an output layer to optimize a coordinate result;

in the hidden layer, a single step data input vector x is added_p(k)＝f(w₁u(k)+w₂x_c(k+1)+b₁) In this case, the output unit vector of the hidden layer is x (k) f (w)₁u(k-1)+w₂x_c(k)+w₃x_p(k)+b₁) In the formula, w₁,w₂,w₃Is a weight, u is an input vector of the neural network, x_cFor hidden layer feedback vectors, b₁F is a sigmoid function, and k is a time;

adding an extended Kalman filter in the output layer of the Elman network to calibrate the original output For the estimated rough three-dimensional trajectory coordinates, the state equation is set as:

set the observation equation toThe Jacobian matrix of the equation of state isIn the above formula, ω_kAngular velocity at time k, η_k,δ_kIs white gaussian noise;

(3.3) the output adopted in the training of the neural network is a three-dimensional coordinate which is collected by an optical dynamic capturing system;

(4) analyzing the posture and track coordinate information of key nodes of the whole body, and evaluating the action quality;

the action quality evaluation method comprises the following specific steps:

(4.1) for a sequence of actions¹O,²O,...,^NO } and the corresponding sequence of the standard action sequence library¹O_s,²O_s,...,^NO_sComparing, and analyzing the action quality;

and sequences in the standard action sequence library, wherein the sequence format of one sensor i is as follows:wherein w_1i,w_2iThe scoring weight is set for the person according to the action type and the sensor part;

(4.2) forⁱO andⁱO_scalculating a standard degree E corresponding to the quality of the motion_iAnd a stability S_iThe specific calculation method is as follows: wherein std represents a standard deviation;

at this time, the process of the present invention,the quality score of the joint whole node to be evaluated is { Q, Q₁,q₂,...,q_NIs totally divided intoWherein q is_min，q_maxRespectively representing the maximum and minimum scores of each part, and the action quality score of each part of the body is q_i＝w_1iE_i+w_2iS_iWherein w is_1i,w_2iCan be based on the type of action (C)₁,C₂) Directly from the standard library.

2. The method of claim 1, wherein the method comprises: step (2.2), the specific bare-handed fitness action for identification comprises the following steps:

3. the method of claim 1, wherein the method comprises: and (4) the first 6 rows of data of the action sequence of the action classification in the step (2.3) are normalized data.

Technical Field

The invention belongs to the field of fitness exercise, and particularly relates to a method for evaluating the quality of fitness actions without instruments based on micro-inertia.

Background

In recent years, the body-building concept is deeply concentrated under the promotion of national fitness policy, and the demand of people on body building is more and more vigorous. With the vigorous development of the internet, the home and on-line body-building method attracts most people to be added into the body-building. However, with the hot tide of body building, the injury caused by improper exercise is also the focus of people's attention.

Lack of scientific-guided body-building activities aiming at physical conditions and easily cause damage to the body. Scientific fitness ways are always sought after. However, since the current gym personal trainer is expensive, and a part of the population chooses to do home autonomous exercises, there is a problem that the standardization of the motion is difficult to estimate. Because the user independently imitates the video action training, lack professional guidance, the action standard nature in the motion process is difficult to judge, can cause permanent damage when serious.

In order to solve the above problems, according to the patent of chinese patent publication (public disclosure) No. CN104888444A, an intelligent glove, method and system for identifying calorie consumption and hand gesture are provided, which can identify fitness data and calorie consumption of a user. This patent utilizes inertial sensor and pressure sensor to carry out the gesture classification to the power training action to can in time warn when the hand gesture takes place to squint. The motion data that this gloves collected the processing is comparatively single, can not carry out effectual aassessment to the whole body gesture of body-building process, also can not carry out aassessment analysis to hand body-building in-process real-time position, leads to the user to the cognitive degree of body-building action overall quality low. Chinese patent publication (publication) No. CN106073793B discloses a posture tracking and identifying method based on micro inertial sensors, which uses wearable micro inertial sensors to collect posture data of a human body and perform posture data processing, tracks and identifies actions and behaviors of a user, and judges the normative degree of the actions of the user according to preset standard actions and provides corresponding correction suggestions. According to the method, the motion process of a person is divided into a plurality of basic actions, the person actions are classified in a template matching mode, the definition of a template library is complex, and position information of each part of the human body in the motion process is lacked in the template library. In evaluating human actions, the normality of user actions is evaluated in a time series distance metric. The method omits the requirement of body building action on each part, only measures the standard degree but neglects the stability, focuses on the integral difference, and ignores the action score of local key position.

The current fitness auxiliary system lacks reasonable overall and local quality assessment means, lacks the judgment of real-time positions of all parts in the human motion process and lacks an efficient classification method for freehand fitness actions in the fitness training process of people. Therefore, it is necessary to guide the fitness training of the user to develop an efficient motion estimation method suitable for the whole motion and considering the local key positions for the free-hand fitness training of the user.

Disclosure of Invention

Aiming at the problem that the existing fitness auxiliary system cannot evaluate the action quality of human fitness, the invention provides a micro-inertia-based non-equipment fitness action quality evaluation method which is used for guiding non-professionals to carry out fitness activities in a healthy, safe and effective manner.

The invention provides a micro-inertia based non-equipment fitness action quality evaluation method, which comprises the following steps:

(1) collecting and processing body-building exercise information;

in the step (1), the information of the body-building exercise is collected and processed, 6-axis inertial sensors are arranged on the human body, N inertial sensors are arranged on the human body, and the number of the inertial sensors is N, wherein the number of the upper body is N_uLower part of the body N_d1 in the waist and abdomen;

(2) a human body fitness state-action two-step classification model;

the two-step classification model of the fitness action without the apparatus comprises the following specific steps:

(2.1) the motion states of the human body are roughly classified by utilizing the information of 3 micro inertial sensors of the upper body, the lower body and the waist and abdomen, wherein the information of the micro inertial sensors with 6 axes is divided into five types, specifically: upper limb movement, lower limb movement, waist and abdomen movement, whole body movement and non-movement state;

(2.2) calling the Elman-Kalman model in the step (3) according to the state classification result to obtain coordinate track information of each sensor, performing fine classification of body-building exercises according to nodes of the whole body, and specifically classifying the exercise state into various actions:

(2.3) for the sensor i, combining the classification result, the action coordinate and the Euler angle to form an action quality evaluation sequence, wherein the specific form is as follows:

wherein x, y, z are three-dimensional coordinates of the attachment part of the motion process sensor, α, β, γ are corresponding Euler angle sequences, C₁The state classification result is evaluated as 1-5, which respectively correspond to 5 motion states, C₂Taking the value of 1-m as an action classification result, respectively corresponding to m actions in each motion state, and updating the value after the action classification step is finished;

(3) an Elman-Kalman trajectory estimation model of human motion;

the Elman-Kalman trajectory estimation model comprises the following specific steps:

(3.1) aiming at the first four motion states of the coarse classification result in the step (2), establishing 4 corresponding motion track estimation models for each sensor;

(3.2) the Elman-Kalman trajectory estimation model structure is specifically that in a hidden layer of an Elman neural network, estimation on future data is added, and an extended Kalman filter is constructed in an output layer to optimize a coordinate result;

in the hidden layer, a single step data input vector x is added_p(k)＝f(w₁u(k)+w₂x_c(k+1)+b₁) In this case, the output unit vector of the hidden layer is x (k) f (w)₁u(k-1)+w₂x_c(k)+w₃x_p(k)+b₁) In the formula, w₁,w₂,w₃Is a weight, u is an input vector of the neural network, x_cFor hidden layer feedback vectors, b₁F is a sigmoid function, and k is a time;

adding an extended Kalman filter in the output layer of the Elman network to calibrate the original output For the estimated rough three-dimensional trajectory coordinates, the state equation is set as:

setting the observation equation to y (k) ═ p (k) + δ_kThe Jacobian matrix of the equation of state isIn the above formula, ω_kAngular velocity at time k, η_k,δ_kIs white gaussian noise;

(3.3) the output adopted in the training of the neural network is a three-dimensional coordinate which is collected by an optical dynamic capturing system;

(4) analyzing the posture and track coordinate information of key nodes of the whole body, and evaluating the action quality;

the action quality evaluation method comprises the following specific steps:

(4.1) for a sequence of actions¹O,²O,...,^NO } and the corresponding sequence of the standard action sequence library¹O_s,²O_s,...,^NO_sComparing, and analyzing the action quality;

and sequences in the standard action sequence library, wherein the sequence format of one sensor i is as follows:wherein w_1i,w_2iThe scoring weight is set for the person according to the action type and the sensor part;

(4.2) forⁱO andⁱO_scalculating a standard degree E corresponding to the quality of the motion_iAnd a stability S_iThe specific calculation method is as follows: wherein std represents a standard deviation;

at this time, the quality score of the joint whole node is { Q, Q ] for the action to be evaluated₁,q₂,...,q_NIs totally divided intoWherein q is_min，q_maxRespectively representing the maximum and minimum scores of each part, and the action quality score of each part of the body is q_i＝w_1iE_i+w_2iS_iWherein w is_1i,w_2iCan be based on the type of action (C)₁,C₂) Directly from the standard library.

As a further improvement of the invention, step (2.2), the specific exercise motions for identification by free-hand include:

as a further improvement of the invention, the first 6 columns of data of the action sequence of the action classification of the step (2.3) are all normalized data.

Compared with the prior art, the invention has the advantages that:

the method of the invention sequentially classifies the motion state and the body-building action of the human body by applying a two-step classification algorithm, and can effectively improve the identification rate of the category of the body-building action of the human body by introducing the estimated motion track coordinate of the human body when the body-building action is classified in the second step. When the human body action quality is evaluated, the nodes of the whole body are subjected to fusion evaluation, so that not only can the whole action quality be fed back, but also the action quality of each part of the body can be evaluated.

The invention provides an Elman-Kalman trajectory estimation algorithm for estimating the motion trajectory of each part of a human body, effectively solves the problems of difficult initial alignment, great environmental influence and the like of the traditional pure inertia calculation, and can be widely applied to the trajectory identification of fixed repeated actions.

Drawings

FIG. 1 is a flow chart of an implementation of a method for evaluating the quality of a non-mechanical fitness action based on micro-inertia;

FIG. 2 is a schematic diagram of a state-action two-step classification algorithm;

FIG. 3 is an Algorithm diagram of the Elman-Kalman trajectory estimation of a single node.

Detailed Description

The invention is described in further detail below with reference to the following detailed description and accompanying drawings:

the invention provides a micro-inertia based fitness action quality evaluation method without equipment, which is used for guiding non-professional persons to carry out fitness activities in a healthy, safe and effective manner.

FIG. 1 is a schematic flow chart of the method of the present invention.

Step S1: collecting and transmitting body-building exercise data. The method specifically comprises the following steps:

s1.1, laying 6-axis inertial sensors at 9 key positions of action postures of a human body to be evaluated, wherein the 6-axis inertial sensors comprise 4 upper limbs (a left forearm, a left upper limb, a right forearm and a right upper limb), 4 lower limbs (a left calf, a left thigh, a right calf and a right thigh) and 1 waist;

s1.2, returning acquired sensor data, triaxial acceleration, triaxial angular velocity and Euler angle, wherein the data acquisition frequency is 50 Hz:

a state-action two-step classification algorithm is shown in fig. 2.

Step S2: the exercise information is used for identifying the body-building action of the human body, and a state-exercise two-step classification algorithm is adopted. The method specifically comprises the following steps:

s2.1, dividing the returned data of each sensor, wherein each two seconds is an action sequence, the size of each sequence is 100 multiplied by 9, one action is,ⁱO′_k＝[a_k ω_k α_k β_k γ_k]including three-axis acceleration, three-axis angular velocity, and euler angles.

S2.2, using machine learning algorithms including, but not limited to, support vector machines, decision trees, random forests, andthe machine learning algorithm of neural network, etc. divides the motion state of human body into 5 classes, which are respectively the upper limb motion, lower limb motion, waist and abdomen motion, whole body motion and non-motion state, and marks C₁The value, the motion sequence extension is 100 x 10, where one behavior,ⁱO′_k＝[a_k ω_kα_k β_k γ_k C₁]the motion state flag is incremented.

S2.3 according to the motion state, namely C₁And calling a corresponding Elman-Kalman motion trail estimation model to estimate the motion trail of the corresponding node. The model is illustrated in a single node.

Fig. 3 is a block diagram of a trajectory estimation model of a single node.

The input of the model is the first 9 columns of the motion sequence, where u (k-1) ═ a_k-1,ω_k-1,α_k-1,β_k-1,γ_k-1) First, add a future one-step action sequence, x, to the hidden layer_p(k)＝f(w₁u(k)+w₂x_c(k+1)+b₁) Therefore, the output of the hidden layer is changed to x (k) ═ f (w)₁u(k-1)+w₂x_c(k)+w₃x_p(k)+b₁) Wherein f is a sigmoid function. And adding extended Kalman filtering in an output layer to further calibrate the estimated value. The moment of the fitness activity can be considered as a uniform acceleration of the movement about the fulcrum, so the equation of state can be set as:wherein the content of the first and second substances,is an estimate of the coordinates having a jacobian matrix ofThe observation equation is y (k) ═ p (k) + δ_k. Wherein eta_k,δ_kAre all gaussian noise. The output is estimated three-dimensional coordinates p (k) ═ x_k y_k z_k)^T. During model training, the true value coordinates are obtained by the optical dynamic capturing system.

S2.4, after the estimated value of the motion trail is obtained, the motion sequence of each sensor is updated to beAnd applying the classification to the action classification of the second step, and after obtaining the classification result, further updating the action sequence to a great curl¹O,²O,...,^NO }, wherein

Step S3: analyzing the posture and track coordinate information of the key nodes of the whole body, and evaluating the action quality, wherein the method specifically comprises the following steps:

s3.1, making a check on the action sequence formed in step 2¹O,²O,...,^NO } according to C₁,C₂Calling a corresponding sequence in a standard action sequence library¹O_s,²O_s,...,^NO_sComparative analysis was performed.

S3.2, analyzing the standard degree and the stability of all the nodes, calculating the standard degree for the node i,calculating the degree of stabilityWhere std is the standard deviation.

S3.3, for the node i, the action local score is as follows: q. q.s_i＝w_1iE_i+w_2iS_iWherein w is_1i,w_2iThe scoring weight is set by human according to the action type and calculated according to C₁,C₂The value is called.

S3.4, Total score for Whole body actionThe final feedback to the user is evaluated as a sequence of scores { Q, Q₁,q₂,...,q_NAnd the exercise quality is measured, and the standard reaching conditions of all parts of the body are measured, so that the user is guided to efficiently and safely carry out fitness activities.

The above description is only one of the preferred embodiments of the present invention, and is not intended to limit the present invention in any way, but any modifications or equivalent variations made in accordance with the technical spirit of the present invention are within the scope of the present invention as claimed.