KR101494779B1 - System and method of human gait phase classification for controlling robots - Google Patents

Abstract

Provided in the present invention is a gait phase classification system classifying a gait phase of a wearer for robots or systems similar to the same, which comprises a sensor unit sensing a ground reaction force (GRF) generated from the ground upon the gait of the wearer; and a controlling unit receiving the ground reaction force from the sensor unit and infers the gait phase state of the wearer by a supervised learning method based on the received ground reaction force to produce information of the gait phase state of the wearer by using the ground reaction force measured in the sensor unit.

Description

Technical Field [0001] The present invention relates to a system and a method for discriminating a walking phase of a wearer for controlling a robot,

The present invention relates to a system and method for classifying human walking steps in a robot or similar system that enhances or assists a wearer ' s walking ability.

Generally, the leg-strength-strengthening robot is a device for reinforcing or assisting the walking ability by worn on the lower body of human being. It is used for various purposes such as rehabilitation medical care, defense and disaster relief. We are researching and developing to help. In order to support human walking ability, appropriate motion for each action is generated or controlled by a robot driver by analyzing human walking motion.

The state of human walking phase is divided into 8 steps in 4 steps for each section. More specifically, human walking is a method of using two legs to move the weight of the body by supporting both the legs and the propulsion. As shown in the figure below, the supporting action in the walking phase is expressed as a stance phase and the propulsion action is represented as a swing phase. A ground reaction force (GRF) is formed at the soles of the soles when the leg contacts the ground during walking. The walking step in which the ground reaction force is formed is observed in all five stages including all stages of the stomach section and pre-stomach section of the stomach section.

Since the characteristics of human walking are different between when the human body is in the stance section and when the human body is in the stance section, the robot operates by dividing the control mode into the angular control mode and the stance control mode. In the stance control mode, the robot is designed to follow the leg motion while generating a large driving force because the robot must be able to withstand the load supported by the wearer and to operate in accordance with the movement of the leg.

On the other hand, in the ramp control mode, the wearer's leg is separated from the ground, so that a large driving force is not required. However, the wearer ' s legs are configured to quickly follow the leg motion as they move for a short time.

The robot uses the walking state information to determine the two control modes. The walking state information necessary for the control mode switching of the robot requires four steps including initial contact, angle of attack, advance angle of attack, and angle of attack.

(1) Initial contact: In the first step, the ground reaction force is observed at the heel of the wearer.

(2) Stance: The ground reaction force is observed at the sole at all stages of the stance section excluding the initial contact.

(3) Pre-swing: As the first step to enter the ramp, the ground reaction force is observed at the tip of the toe.

(4) Swing: Ground reaction force is not observed at all stages of the ramp section excluding the pre-ramp.

As described above with respect to the walking state of a human being, it is divided into a stance, which is a state in which the walking operation is largely divided, and a swing, which is a step of forwarding the leg forward. During the effect of walking, the legs support the load of the human being in the stance state, and when the stance is carried out, the legs are moved forward and the forward is propelled. Accordingly, the robot performs a different control command for each walking behavior, and a walking step classification system is required to determine the control command.

In the gait step classification system according to the related art, generally, a sensor is installed on the bottom of the robot or on the foot surface of the wearer of the robot, and the gait phase is divided by the ground reaction force generated when the leg contacts the ground . In addition, since the conventional walking step classification system directly converts a sensor value into a walking state based on a threshold value set arbitrarily in order to distinguish each walking step, erroneous information can be provided according to the observed sensor value.

For example, FIG. 1 is a graph showing data obtained from a sensor measuring a ground reaction force. Referring to FIG. 1, the ground reaction force information obtained from the sensor varies depending on the human body characteristic of the wearer, and the noise and the like are generated in the indoor and outdoor environments in which the robot is operated, It may be difficult to provide accurate information if you simply use thresholds or use established information depending on the combination. There is also a threshold optimization problem to determine the appropriate threshold for the situation.

Therefore, development of a system and method for providing more precise information on the walking step of the wearer while overcoming the characteristics of the sensor can be considered.

The present invention is to provide a gait phase classification system and method that provides reliable gait phase information without using a threshold optimization process, while using ground reaction forces including noise due to wearer or environmental changes.

According to another aspect of the present invention, there is provided a walking step classification system including a sensor unit for measuring a ground reaction force (GRF) generated from the ground according to the wearer's walking, And a control unit for controlling the wearer so as to generate a gait step state information of the wearer using the ground reaction force measured by the gait measuring unit, And a control unit configured to infer by the method.

According to an example of the present invention, the controller may be configured to calculate and compare conditional probabilities for each of the walking state states of the wearer based on the ground reaction force.

 Wherein the control unit comprises: a sensor data processing unit for extracting feature data from the ground reaction force received from the sensor unit; and a control unit for calculating the feature data transmitted from the sensor data processing unit using a Naïve Bayesian technique, And a gait step classification generating unit configured to generate the gait information.

Wherein the control unit is configured to learn the walking step state information using the ground reaction force measured by the sensor unit and transmit the parameter obtained by the learning to the walking step classification generating unit so as to provide the walking step state reference of the wearer And a gait classification model learning unit.

According to another embodiment of the present invention, the sensor unit is disposed so that at least a part of the sensor unit is pressed between the ground and the wearer during walking of the wearer, and the ground surface reaction force is measured from a compressive force varying with the change of the wearer's gait A load cell may be included.

According to another embodiment of the present invention, the sensor unit includes a piezoelectric element arranged to press at least a part between the ground and the wearer on walking of the wearer, and generating a voltage by mechanical deformation, The element may be configured to measure the ground reaction force from a change in voltage generated in accordance with the wearer's walking.

In order to realize the above-mentioned problem, the present invention proposes a gait step classification method. The gait step classification method comprises the steps of measuring a ground reaction force (GRF) generated from the ground according to the walking of the wearer with a sensor unit, And a gait step state of the wearer based on the ground reaction force is derived by a supervised learning method so as to generate gait step state information of the wearer using the measured ground reaction force, .

According to an embodiment of the present invention, the step of discriminating the walking step of the wearer may include the step of calculating the walking step state information of the wearer using the measured ground surface reaction force, Calculating a conditional probability for each of the plurality of conditional probabilities, and comparing the calculated conditional probabilities with each other and determining the walking phase state having the largest value among the conditional probabilities as the walking phase of the wearer .

Wherein the sensor unit includes a plurality of measurement points disposed at different positions and the step of calculating the conditional probability is performed to calculate a conditional probability for each of the walking step states based on the ground reaction force measured at the plurality of measurement points .

The conditional probability can be obtained by the following equation using the Naïve Bayesian technique.

)과 상기 지면반력의 변화량(

)이다. 또한, p(c|x)는 사후확률, p(c)는 보행단계상태 c의 사전확률, p(x_i|c)는 보행단계상태 c에 대한 각각의 특징변수 x_i의 우도확률이다. 또한, 전체 특징변수 x에 대한 모집단의 확률 p(x)는 상수로 표현 가능하므로 생략한다.Here, c is the gait step state, and x (characteristic variable) is the ground reaction force

) And a variation amount of the ground surface reaction force

)to be. Further, p (c | x) is the posterior probability, p (c) is a priori probability of the walking phase state c, p | a (x _i c) is the likelihood probability of each feature parameter x of the walking phase state c _i. In addition, the probability p (x) of the population for the entire feature variable x can be represented by a constant and is omitted.

The step of calculating the conditional probability may include extracting the feature data from the ground reaction force transmitted from the sensor unit, and generating the walking state information of the wearer using the feature data.

The specific variable x may be the feature data obtained from the extraction of the feature data.

According to another embodiment of the present invention, the discriminating step of the wearer is a step of discriminating the walking step from the walking segment model learning part, which is configured to learn the walking step state information using the ground reaction force measured by the sensor part, And to determine the walking step of the wearer by receiving the parameters obtained by the learning.

A method of dividing a walking step of a wearer for controlling a robot according to the present invention and an effect of the system will be described as follows.

According to at least one of the embodiments of the present invention, the conditional probability for each walking state of the robot wearer is calculated using the Naïve Bayesian classification technique, which is one of concrete methods of the map learning classification technique, and the calculated conditional probability And a value having the greatest probability value among them is discriminated as the present walking step of the wearer.

Accordingly, it is possible to improve reliability by a stochastic approach without constructing a method for optimizing a separate threshold value in various situations in which noises occur when the walking step is divided.

Further, according to the present invention, the walking step can be divided into four steps according to the purpose of the system, and the steps below or above the four steps can be switched through the data learning process.

1 is a graph showing data obtained from a human body sensor measuring a ground reaction force;
2 is a conceptual diagram illustrating a configuration of a walking stage classification system according to an embodiment of the present invention;
3 is a conceptual diagram showing a flow of a gait step classification method by the gait step classification system shown in Fig.

Hereinafter, a method of dividing a walking step of a wearer for controlling the robot according to the present invention and a system thereof will be described in detail with reference to the accompanying drawings.

In the present specification, the same or similar reference numerals are given to different embodiments in the same or similar configurations. As used herein, the singular forms "a", "an" and "the" include plural referents unless the context clearly dictates otherwise.

2 is a conceptual diagram showing a configuration of a walking stage classification system 100 according to an embodiment of the present invention.

Referring to FIG. 2, in the system for classifying the walking step of the wearer for the robot control, the walking step classification system 100 includes a sensor unit 110 and a control unit 120.

The sensor unit 110 measures the ground reaction force (GRF) generated from the ground during walking of the wearer wearing the robot. For example, the sensor unit 110 may include a load cell (not shown). The load cell may be arranged to press at least part of the space between the ground and the wearer when the wearer walks the robot, and may measure the ground reaction force from a compressive force depending on the change of the wearer's gait. The plurality of load cells may be provided.

In addition, the sensor unit 110 may include a piezoelectric element (not shown).

The piezoelectric element is arranged so that at least part of the piezoelectric element is pressed between the ground and the wearer during walking of the wearer, and generates a voltage when mechanical deformation occurs. The piezoelectric element may be configured to measure the ground reaction force from a change in the voltage generated in accordance with the wearer's walking. The plurality of piezoelectric elements may be provided.

The control unit 120 predicts the walking state of the wearer based on the ground reaction force measured by the sensor unit 110 by a supervised learning method. Here, the map learning is a method of machine learning for estimating one associated result value from the ground reaction force, that is, training data.

More specifically, the control unit 120 receives the ground reaction force from the sensor unit 110, and calculates and compares conditional probabilities for each of the walking condition states of the wearer on the basis of the received ground reaction force. Here, the conditional probability means a probability of occurrence of the state B, for example, when the state A occurs in any two states A and B. [

1, the walking step sorting system 100 includes an interface for connecting the sensor unit 110 and the control unit 120 so as to transmit and receive data between the sensor unit 110 and the control unit 120, (10). ≪ / RTI >

The control unit 120 may include a sensor data processing unit 130 and a walking step classification generating unit 140.

The sensor data processing unit 130 extracts the feature data from the ground reaction force transmitted from the sensor unit 110. For example, the feature data may include a change amount of the ground reaction force or the ground reaction force measured by the sensor unit 110.

The gait phase classification generator 140 generates the gait phase state information of the wearer after calculating the feature data transmitted from the sensor data processor 130 using the Naïve Bayesian method. The Naïve Bayesian method is a method for calculating the conditional probability, and a description thereof will be given below with reference to a description of a walking step classification method.

In addition, the control unit 120 may further include a walking segment model learning unit 150. [

3, the gait classification model learning unit 150 is configured to learn the gait phase state information using the ground reaction force measured by the sensor unit 110, To the walking-phase classifying and generating unit 140, the parameters obtained by the learning. Accordingly, the gait phase classification generator 140 can generate more reliable gait phase state information using the parameters generated from the learned data.

Hereinafter, a gait phase separation method by the gait phase classification system 100 will be described with reference to Fig.

FIG. 3 is a conceptual diagram showing a flow of a gait step classification method by the gait step classification system 100 shown in FIG.

3, the gait step classification method includes the steps of measuring a ground reaction force (GRF) generated from the ground according to the walking of the wearer with the sensor unit 110, And a step of discriminating the walking step of the wearer by analogizing the state by a supervised learning method.

More specifically, the step of discriminating the walking step of the wearer may further comprise the step of calculating the walking step state information of the wearer using the measured ground surface reaction force, Calculating a probability and comparing the calculated conditional probabilities with each other and discriminating the walking step state having the largest value among the conditional probabilities as the walking step of the wearer.

The above figure is a conceptual diagram showing the control mode transition operation according to the walking state. As shown in the figure above, the ramp control mode is activated when the walking step is in the upright position and when the walking step is in the upright position. Also, in the case of the initial contact or preheating, the control mode state is transitioned from the ramp control mode to the angular control mode or from the angular control mode to the ramp control mode.

If the control mode state is abruptly changed, discontinuity may occur in the control input, which may act as a resistance to the movement of the wearer, so that the control input continuity may be considered.

The step of calculating the conditional probability may include extracting the feature data from the ground reaction force transmitted from the sensor unit 110 and generating the walking state information of the wearer using the feature data can do.

The sensor unit 110 may include a plurality of measurement points disposed at different positions, and the step of calculating the conditional probability may include calculating a conditional probability for each of the walking step states on the basis of the ground reaction force measured at the plurality of measurement points Can be made to calculate the probability. For example, referring to the following figure, the sensor unit 110 detects the heel (x ₁ ), the ball outside the ball (x ₂ ), the ball inside the ball (x ₃ ) and the toe (x ₄ ) And four measurement points made up of two points. At this time, the plurality of measurement points are not necessarily limited to four measurement points, and may be constituted by a plurality of measurement points composed of an arbitrary number to measure the wearer's ground surface reaction force.

Hereinafter, the calculation of the conditional probability using the Naive Bayesian method will be described in detail.

별로 속할 조건부 확률을 각각 계산하고 그 중 가장 큰 확률 값을 나타내는 보행단계상태를 현재의 보행단계상태 C_gaitphase로 결정한다.The method of stochastic classification of the gait phase state is as follows. When the feature variable x is given as shown in the following equation (1)

The conditional probabilities belonging to each one are calculated, and the state of the gait step indicating the largest probability value is determined as the current gait phase state C _gaitphase .

식(1)

Equation (1)

The conditional probability can be obtained by the following equation (2) using the Naïve Bayesian technique.

식(2)

Equation (2)

)과 상기 지면반력의 변화량(

)이다. 또한, p(c|x)는 사후확률, p(c)는 보행단계상태 c의 사전확률, p(x_i|c)는 보행단계상태 c에 대한 각각의 특징변수 x_i의 우도확률이다.Here, c is the gait step state, and x (characteristic variable) is the ground reaction force

) And a variation amount of the ground surface reaction force

)to be. Further, p (c | x) is the posterior probability, p (c) is a priori probability of the walking phase state c, p | a (x _i c) is the likelihood probability of each feature parameter x of the walking phase state c _i.

In addition, the probability p (x) of the population for the entire feature variable x can be represented by a constant and is omitted.

Further, the specific variable x may be the feature data obtained from the extraction of the feature data.

In the Naive Bayesian method, the feature variables x _i are assumed to be conditionally independent of each other. In addition, there is a problem that the actual calculated probability appears to be much smaller than expected.

However, there is a merit that the probability calculation process is simple and the calculation cost is low. In the case where the robot distinguishes the walking step, it is necessary to be able to distinguish the walking step at a high speed on the system rather than the probability value. Therefore, the Naive Bayesian method has an advantage of low computational cost in terms of calculation speed.

?의 우도함수를 나타내는데 사용할 수 있다. 또한, 커널밀도분포모델은 나이브 베이지안 기법의 변종인 플랙시블 베이지안 기법에서 처음 제시되었다. 커널밀도분포 모델은 정규분포 대비 지역적인 특성을 잘 표현하기 때문에 특징 변수

의 우도함수를 나타내는데 사용할 수 있다.The likelihood probability p (x _i | c) for each feature variable is constructed by using the normal distribution and the kernel density distribution together. The normal distribution model is mainly used in the Naive Bayesian method and because it expresses the global characteristics well,

Can be used to represent the likelihood function of?. In addition, the kernel density distribution model was first presented in a flexible Bayesian technique, a variant of the Naive Bayesian technique. Since the kernel density distribution model expresses the regional characteristics versus the normal distribution well,

Can be used to represent the likelihood function of.

The likelihood probability model for the gait phase classification is shown in Equation (3) below.

식(3)

Equation (3)

의

와

는 각 특징변수 집합 x_i의 평균과 분산을 말하며, 커널밀도분포를 갖는 확률밀도함수

에서 n은 특징변수 집합 내 원소의 수,

은 특징변수의 원소 값

,?

는

이며

는 보행단계상태 c에 속하는 특징변수들의 모든 원소의 수

이다.Here, a probability density function having a normal distribution

of

Wow

Is the mean and variance of each feature set x _i , and is the probability density function with kernel density distribution

Where n is the number of elements in the feature variable set,

Is the element value of the characteristic variable

,?

The

And

The number of all the elements of the characteristic variables belonging to the walking phase state c

to be.

,?

,?

,?

는 보행단계를 미리 구분한 훈련데이터집합과 최대우도추정기법(Maximum Likelihood Estimation)과 커널밀도추정(Kernel Density Estimation)에 의해 결정된다.The parameters of the prior probability p (c) and likelihood probability p (x | c) in the walking step classification algorithm

,?

,?

,?

Is determined by the training data set which classifies the walking steps in advance and the maximum likelihood estimation method and the kernel density estimation method.

On the other hand, the discriminating step of the wearer is a step of discriminating the wearer's walking step from the walking segment model learning unit 150 which is configured to learn the walking step state information using the ground reaction force measured by the sensor unit 110, And a step of determining a walking step of the wearer by receiving the parameter.

According to the system and method of the present invention described with reference to FIGS. 2 and 3, in generating information on the walking state of the wearer, it is possible to calculate the most probable walking Provide step status. Accordingly, it is possible to reduce the occurrence of errors in the step-by-step state classification caused by wearing a plurality of wearers having different gait patterns by using one robot, and when the robot is operated on the ground having various shapes, 110 can be provided with reliable gait phase state information even when the noise of the measured data is increased.

Meanwhile, in the system and method of the present invention, the gait step state is described based on the four-step gait method consisting of the initial gait, the gait, the preliminary gait, and the ramp. However, But may be divided into two stages, three stages or eight stages through the learning process of the supervised learning method.

However, the scope of the present invention is not limited to the configuration and method of the embodiments described above, and all or some of the embodiments may be selectively combined so that various modifications may be made to the embodiments. In addition, the present invention can be applied to all equivalents of inventions, such as inventions that can be modified, added, deleted, or replaced at the level of those skilled in the art, It belongs to the scope is self-evident.

100: walking step classification system 110:
120: control unit 130: sensor data processing unit
140: walking step classifying section 150: walking classifying section learning section

Claims (16)

A system for classifying a walking step of a wearer wearing a robot or a system for reinforcing or assisting a walking ability of a wearer,
A sensor unit for measuring a ground reaction force (GRF) generated from the ground according to the wearer's walking; And
Receiving the ground reaction force from the sensor unit so as to generate walking condition information of the wearer using the ground reaction force measured by the sensor unit and receiving the walking reaction state of the wearer with reference to the received ground reaction force, And a control unit configured to infer by a supervised learning method,
Wherein the control unit controls the wearer using the navy Bayesian technique for each of the walking state states of the wearer based on the ground reaction force,

And calculating a conditional probability according to the following equation: < EMI ID = 1.0 > delete The method according to claim 1,
Wherein,
A sensor data processing unit for extracting characteristic data from the ground reaction force transmitted from the sensor unit; And
And a gait step classification section for calculating the gait step state information of the wearer after calculating the feature data transmitted from the sensor data processing section using a Naïve Bayesian method. The method of claim 3,
Wherein,
And a gait classification module that receives the gait phase state information of the wearer using the ground reaction force measured by the sensor unit and transmits the parameter obtained by the learning to the gait step classification generator, And a learning unit. 5. The method of claim 4,
And a walking step classifying unit configured to receive the parameters obtained by learning from the learning unit and to generate walking state information using the parameters generated from the learned data. The method according to claim 1,
The sensor unit includes:
And a load cell arranged to press at least part of the space between the ground and the wearer when walking the wearer, the load cell measuring the ground reaction force from a compressive force varying with the change of the wearer's gait. The method according to claim 1,
The sensor unit includes:
And a piezoelectric element arranged to be pressed at least partly between the ground and the wearer when walking on the wearer and generating a voltage by mechanical deformation,
Wherein the piezoelectric element is configured to measure the ground reaction force from a change in voltage caused by the wearer's walking. A method for distinguishing a walking step of a wearer wearing a robot or a system that enhances or assists a wearer's walking ability,
Measuring a ground reaction force (GRF) generated from the ground according to the wearer's walking with the sensor unit; And
And a gait step state of the wearer on the basis of the ground reaction force is derived by a supervised learning method so as to generate gait step state information of the wearer using the measured ground reaction force, , ≪ / RTI >
The step of discriminating the walking step of the wearer comprises:
Using the measured knee bending method for each of the walking step states on the basis of the measured floor reaction force so as to generate the walking state information of the wearer using the measured floor reaction force,

With a conditional probability following the equation: < RTI ID = 0.0 > And
Comparing the calculated conditional probabilities with each other, and discriminating the walking step state having the largest one of the conditional probabilities as the walking step of the wearer. delete 9. The method of claim 8,
Wherein the sensor unit includes a plurality of measurement points disposed at different positions,
Wherein the step of calculating the conditional probability is performed to calculate a conditional probability for each of the walking step states on the basis of the ground reaction force measured at the plurality of measurement points. 11. The method of claim 10,
Wherein the sensor unit includes a plurality of measurement points disposed at different positions,
Wherein the step of calculating the conditional probability is performed to calculate a conditional probability for each of the walking step states on the basis of the ground reaction force measured at the plurality of measurement points. 9. The method of claim 8,
The above conditional probability is obtained by using the Naive Bayesian method,
c is the gait phase state, and the feature variable x is the ground reaction force

) And a variation amount of the ground surface reaction force

), P (c | x) is the posterior probability, p (c) is the prior probability of the walking step state c, and p (xi | c) is the likelihood probability of each feature variable x A step of discriminating a walking step. 13. The method of claim 12,
Wherein the step of calculating the conditional probability comprises:
Extracting feature data from the ground reaction force transmitted from the sensor unit; And
And generating walking state information of the wearer using the feature data. 14. The method of claim 13,
Wherein the feature variable x is the feature data obtained from the extracting step. 14. The method of claim 13,
A method for classifying the wearer's walking state,
Given the feature variable x, each gait phase state

Calculating a conditional probability to belong to each of the plurality of walking steps, and determining a walking step state indicating the largest probability value among the probability values according to the following equation as the current walking step state.

9. The method of claim 8,
The step of discriminating the wearer as a walking step comprises:
And a step of discriminating a walking step of the wearer by receiving a parameter obtained by the learning from a walking segment model learning unit configured to learn the walking step state information using the ground reaction force measured by the sensor unit.

Images