Human body upper limb thrown object modeling method based on hidden Markov model

Abstract

The invention discloses a human body upper limb thrown object modeling method based on a hidden Markov model. The method comprises the following steps: S1, recording multiple groups of upper limb action original data when a human body throws an object by adopting a motion capture system; s2, processing the original action data, and extracting a shoulder joint angle, an elbow joint angle and a wrist joint angle in a throwing process as action characteristics, namely joint angle vectors; s3, determining the parameter dimension of the hidden Markov model and carrying out initialization; s4, training a hidden Markov model by using the joint angle vector, and carrying out iterative training until convergence to obtain an upper limb action hidden Markov model when the human body throws the object; and S5, outputting an action sequence by using the upper limb action model when the human body throws the object. Modeling is carried out on the action of throwing the object by the human body by applying the hidden Markov model, the action sequence with the action characteristics of throwing the object by the human body can be output in batches, and the learning efficiency of the robot and the adaptability and generalization of throwing the object by the robot can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of human motion modeling, and particularly relates to a method for modeling the object throwing of the human upper limb based on the hidden Markov model. Background Art

[0002] With the rapid development of technology, robots are increasingly widely used in daily life. People expect robots to master human skills to assist humans in completing complex tasks. However, currently, robots still face major challenges in learning and executing human skills. For example, although the basic human skill of throwing is relatively simple for ordinary people, it poses extremely high technical requirements for robots.

[0003] In the prior art, researchers have conducted various studies and explorations to enable robots to master the throwing skill. For example, establishing physical models, inverse kinematic modeling, introducing visual feedback, designing special mechanisms, etc. Although certain achievements have been made in the research of robot throwing skills in the prior art, there are still the following problems and deficiencies:

[0004] (1) Insufficient throwing accuracy and adaptability. Most studies focus on throwing under specific targets or conditions, and the adaptability to multiple targets and multiple environments is limited;

[0005] (2) Complexity of motion control. The robot throwing action involves complex kinematic and dynamic analyses, and the prior art cannot fully simulate the flexible throwing motion of humans;

[0006] (3) Low learning efficiency. Robots have a slow learning speed and weak generalization ability for throwing skills, and it is difficult to apply them to actual scenarios. Summary of the Invention

[0007] The main purpose of the present invention is to overcome the drawbacks and deficiencies of the prior art, and propose a method for modeling the object throwing of the human upper limb based on the hidden Markov model.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] A method for modeling the object throwing of the human upper limb based on the hidden Markov model, comprising the following steps:

[0010] S1. Using a motion capture system to record multiple groups of original motion data of the upper limb when a human throws an object;

[0011] S2. Processing the collected original motion data, and extracting the shoulder joint angle, elbow joint angle, and wrist joint angle during the throwing process as motion features, namely the joint angle vector;

[0012] S3. Determining the parameter dimension of the hidden Markov model and initializing it;

[0013] S4. Train a hidden Markov model using the joint angle vectors, and iterate the training until convergence to obtain the upper limb movement hidden Markov model of the human body when throwing an object.

[0014] S5. Use the upper limb movement model of the human body when throwing an object to output an action sequence.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0016] 1. The method of the present invention uses a hidden Markov model to model the movement of the human body throwing an object. The model can batch output action sequences with the action characteristics of the human body throwing an object, which can improve the learning efficiency of the robot and enhance the adaptability and generalization of the robot throwing an object. The hidden Markov model is a Bayesian-based probability model. One of its major features is that it can learn the hidden features of the training data. For this method, the hidden Markov model can be used to learn the characteristics of the movement when the human body throws an object, that is, regardless of the weight of the object thrown by the human body and the distance of the throw, the action sequences output by the model established by this method have certain action characteristics of the throwing process, and have greater generalization compared with the prior art. Secondly, through this model, the action sequences of the human body throwing an object can be batch output, improving the subsequent learning efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flowchart of the method of the present invention;

[0018] Figure 2 is a schematic diagram of pasting marker points on the upper limb of the human body in the embodiment;

[0019] Figure 3 is a visual representation of the shoulder joint angle, elbow joint angle, and wrist joint angle during the process of throwing an object;

[0020] Figure 4 is a schematic diagram of the model outputting an action sequence in the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following further describes the present invention in detail with reference to the embodiments and the drawings, but the embodiments of the present invention are not limited thereto.

[0022] Embodiment

[0023] As Figure 1 shown, the present invention, a method for modeling the upper limb of a human body throwing an object based on a hidden Markov model, includes the following steps:

[0024] S1. Use a motion capture system to record multiple groups of original data of the upper limb movements of the human body when throwing an object; in this embodiment, specifically including:

[0025] As shown Figure 2 in the figure, multiple marker points are pasted on the upper limb of the human body, denoted as points A, B, C, D, E, and F respectively; among them, as Figure 2 shown in the figure, point B and point D are pasted at the elbow joint and the wrist joint respectively, point A is pasted on the upper arm, point C is pasted on the lower arm, and points E and F are pasted on the palm;

[0026] The motion capture system is used to record the original motion data of the upper limb of the human body when throwing an object. Specifically, the experimental personnel sit upright on a chair and throw sandbags of different masses to positions at different distances, and the motion capture system records the spatial Cartesian coordinate system coordinates of each marker point.

[0027] S2. Process the collected original motion data, and extract the shoulder joint angle, elbow joint angle, and wrist joint angle during the throwing process as motion features;

[0028] In this embodiment, specifically:

[0029] Using the transformation relationship, the spatial position coordinates of the marker points are converted into the joint angle information of the upper limb of the human body. Denote the spatial three-dimensional coordinates of any marker point P at time t as P t (x P,t , y P,t , z P,t ). According to points A and B, determine the shoulder joint angle θ s,t , and the formula is:

[0030]

[0031] where x A,t , y A,t , and z A,t are the x-axis coordinate, y-axis coordinate, and z-axis coordinate of point A respectively; x B,t , y B,t , and z B,t are the x-axis coordinate, y-axis coordinate, and z-axis coordinate of point B respectively;

[0032] According to points A, B, and C, determine the shoulder joint angle θ s,t , and the formula is:

[0033]

[0034] where x C,t , y C,t , and z C,t are the x-axis coordinate, y-axis coordinate, and z-axis coordinate of point C respectively;

[0035] According to points D, E, and F, determine the palm center coordinate O, and the formula is:

[0036] O = (D t + E t + F t ) / 3

[0037] where D t , E t , F t are the three - dimensional space coordinates of points D, E, and F respectively. When using a motion capture system to collect the motion state of the palm part, it is easy to lose data points due to self - occlusion. Therefore, multiple points are used to determine the center coordinate O of the palm rigid body.

[0038] Furthermore, the wrist joint angle θ is determined based on points C, D, and O w,t , and the formula is:

[0039]

[0040] where x D,t , y D,t , and z D,t are the x - axis coordinate, y - axis coordinate, and z - axis coordinate of point D respectively; x O,t , y O,t , and z O,t are the x - axis coordinate, y - axis coordinate, and z - axis coordinate of the center coordinate O of the palm rigid body respectively;

[0041] During the process T of throwing an object, the shoulder joint angle, elbow joint angle, and wrist joint angle are extracted as the action features of this process, that is, the joint angle vector, which is expressed as:

[0042] x l (t) = {θ s,t , θ e,t , θ s,t}, 1 ≤ t ≤ T.

[0043] As Figure 3 shown, it is a visual representation of the shoulder joint angle, elbow joint angle, and wrist joint angle during the process of throwing an object.

[0044] S3. Determine the parameter dimension of the hidden Markov model and initialize it;

[0045] The core idea of the hidden Markov model is that its state itself cannot be directly observed, but can be indirectly inferred through the observed vector sequence. Each observed vector corresponds to a certain state and is generated with a certain probability density distribution. The hidden Markov model consists of two random processes: one is the random transition of the state sequence, and the other is the probability correlation between the state and the observed vector. In step S3, the hidden Markov model is represented by the available state set (N), the observation set (M), the initial state probability (Π), the state transition probability matrix (A), and the output probability matrix (B);

[0046] In step S3, the parameter dimensions of the available state set (N), the observation set (M), the initial state probability (Π), the state transition probability matrix (A), and the output probability matrix (B) are determined according to the actual situation and randomly initialized.

[0047] In this embodiment, the available state set (N) is determined. The available state set represents the set of hidden states in a hidden Markov model. As Figure 3 shown, frames 20-40 are the key processes. After determining the key processes during the throwing process, the highest and lowest points of the shoulder joint angle, the lowest and highest points of the elbow joint angle, and the highest point of the wrist joint angle are selected. These 5 key points are recorded as the 5 states of the key process, and it is determined that the number of elements in the available state set (N) is 5. Furthermore, it is determined that the dimensions of the initial state probability (Π) and the state transition probability matrix (A) are both 5-dimensional.

[0048] The observation set (M) is determined. The initial state probability represents the set of directly observable data in a hidden Markov model, which can be divided into discrete data or continuous data; the x l (t) obtained in step S2 is the element in the observation set, that is, the angle information of the three joints at each moment. Since it is not discrete observation data, the subsequent output probability matrix (B) is represented by the Gaussian distribution probability density function b j (x).

[0049] The initial state probability (Π) is determined. The initial state probability represents the probability of each state appearing at the beginning. Assuming that the probabilities of the five states of the key process are equal, that is

[0050] Π init =(0.2, 0.2, 0.2, 0.2, 0.2).

[0051] The state transition probability matrix (A) is determined. The state transition probability matrix represents the probability of transition between each state. For example, a ij represents the probability of state i transitioning to state j; assuming that a certain state can only remain in its original state or transition to the next state, and the probabilities of the two are equal, then there is:

[0052]

[0053] The output probability matrix (B) is determined. The output probability matrix represents the probability of the observed value corresponding to each state. Usually, a single Gaussian distribution probability density function is not sufficient to completely describe a process. Therefore, an output probability matrix composed of three Gaussian distribution probability density functions is adopted:

[0054]

[0055] Among them, cjm The coefficient representing the m-th Gaussian component can be initialized as:

[0056] c j,init =(0.3, 0.3, 0.4)

[0057] For the Gaussian distribution probability density function, we have:

[0058]

[0059] μ j,init =((1, 1, 1), (1, 1, 1), (1, 1, 1))

[0060]

[0061] Finally, the determination and initialization of the Hidden Markov Model parameter dimensions are completed.

[0062] S4. Use the joint angle vector x l (t) to train the Hidden Markov Model, and iterate the training until convergence to obtain the upper limb motion model when the human body throws an object, which can batch output the motion sequences of the human upper limb when throwing an object; in this embodiment, specifically:

[0063] By the forward algorithm, assuming that the forward observation values are known, the probability that it belongs to state i is:

[0064] α t (i) = P(o 1 , o 2 , …, o t , s t = i|λ)

[0065] Then we have:

[0066]

[0067] Then by the backward algorithm, assuming that the backward observation values are known, the probability that it belongs to state i is:

[0068] β t (i) = P(o t+1 , o t+2 , …, o T , s t = i|λ)

[0069] Then we have:

[0070]

[0071] Applying the Baum-Welch algorithm, given the complete observation sequence o, the probability that the observation value o t belongs to state i is:

[0072]

[0073] Given a complete observation sequence o, if the state at time t is i, then the probability that the state at time t + 1 is j is:

[0074]

[0075] Then the final iteration formula is determined as:

[0076]

[0077]

[0078] The iteration termination condition is:

[0079] |P(o|λ i+1 ) - P(o|λ i )| < 0.001 The final trained model is:

[0080] Π final = (1.0, 0.0, 0.0, 0.0, 0.0)

[0081]

[0082]

[0083] S5. Use the upper limb motion model when a human throws an object to output a motion sequence; as Figure 4 shown, it can be seen that it has the characteristics of a human upper limb throwing an object and can output in batches.

[0084] It should also be noted that in this specification, terms such as "including", "comprising" or any other variation thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article or device including the said element.

[0085] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for modeling objects thrown by human upper limbs based on hidden Markov model, characterized in that: The following steps are involved: S1. Use a motion capture system to record multiple groups of raw data of upper limb movements when a person throws an object; S2, processing the collected raw action data, extracting the shoulder joint angle, elbow joint angle and wrist joint angle in the throwing process as action features, i.e., joint angle vector; S3, determine the dimension of hidden Markov model parameters and initialize them; S4, using the joint angle vector to train the hidden Markov model, iteratively training until convergence, to obtain the hidden Markov model of the upper limb action when the human body throws the object; S5. Output an action sequence using the upper limb action model of a human body when throwing an object.

2. A method for modeling human upper limb projectiles based on a hidden Markov model according to claim 1, characterized in that: Step S1 specifically includes: Six marker points are pasted on the upper limbs of the human body, which are marked as points A, B, C, D, E and F. Point B and point D are pasted on the elbow joint and wrist joint respectively, point A is pasted on the upper arm, point C is pasted on the lower arm, and points E and F are pasted on the palm. The motion capture system is used to record the original data of the human upper limbs throwing objects. Specifically, the experimenter sits on a chair and throws sandbags of different masses to positions at different distances. The motion capture system records the spatial Cartesian coordinate system coordinates of each marker point.

3. A method for modeling human upper limb projectiles based on a hidden Markov model according to claim 2, characterized in that: Step S2 is specifically as follows: Using the transformation relationship, the spatial position coordinates of the marker point are converted into the joint angle information of the upper limb of the human body. The spatial three-dimensional coordinates of any marker point P at time t are recorded as P t (x P,t ,y P,t ,z P,t ), determine the shoulder joint angle θ based on points A and B s,t , the formula is: Among them, x A,t ,y A,t and z A,t are the x-axis coordinate, y-axis coordinate, and z-axis coordinate of point A respectively; B,t ,y B,t and z B,t are the x-axis coordinate, y-axis coordinate and z-axis coordinate of point B respectively; Determine the shoulder joint angle θ based on points A, B, and C S,t , the formula is: Among them, x C,t ,t C,t and z C,t are the x-axis coordinate, y-axis coordinate, and z-axis coordinate of point C respectively; Determine the center coordinate O of the palm rigid body based on points D, E, and F. The formula is: O=(D t +E t +F t ) / 3 Among them, D t , E t , F t are the three-dimensional coordinates of points D, E and F respectively; Then determine the wrist joint angle θ based on points C, D, and O w,t , the formula is: Among them, x D,t ,y D,t and z D,t are the x-axis coordinate, y-axis coordinate, and z-axis coordinate of point D respectively; O,t ,y O,t and z O,t They are the x-axis coordinate, y-axis coordinate and z-axis coordinate of the palm rigid body center coordinate O respectively; In the object throwing process T, the shoulder joint angle, elbow joint angle and wrist joint angle are extracted as the action features of the process, that is, the joint angle vector, which is expressed as: x l (t)={θ s,t ,i e,t ,i s,t },1≤t≤T.

4. The method for modeling a human upper limb projectile based on a hidden Markov model according to claim 3, characterized in that: In step S3, the hidden Markov model is represented by an available state set, an observation set, an initial state probability, a state transition probability matrix, and an output probability matrix; Step S3 specifically determines the parameter dimensions of the available state set, observation set, initial state probability, state transition probability matrix and output probability matrix according to actual conditions and initializes them randomly.

5. The method for modeling a human upper limb projectile based on a hidden Markov model according to claim 4, characterized in that: Determine the available state set, which represents a hidden state set in a hidden Markov model; after determining the key process in the throwing process, select the highest and lowest points of the shoulder joint angle, the lowest and highest points of the elbow joint angle, and the highest point of the wrist joint angle, and record the five key points as the five states of the key process, determine the number of elements in the available state set, and then determine the initial state probability dimension and the state transition probability matrix dimension.

6. A method for modeling human upper limb projectiles based on a hidden Markov model according to claim 5, characterized in that: Determine the observation set. The initial state probability represents a set of directly observable data in a hidden Markov model, which is divided into discrete data or continuous data. The x obtained in step S2 l (t) is the element in the observation set, that is, the angle information of the three joints at each moment.

7. The method for modeling a human upper limb projectile based on a hidden Markov model according to claim 6, characterized in that: Determine the initial state probability. The initial state probability indicates the probability of each state appearing at the beginning. Assume that the probabilities of the five states of the key process are equal, that is, P init =(0.2,0.2,0.2,0.2,0.2).

8. The method for modeling a human upper limb projectile object based on a hidden Markov model according to claim 7, characterized in that: Determine the state transition probability matrix. The state transition probability matrix represents the probability of transition between states. Assuming that a state can only maintain the original state or transfer to the next state, and the probabilities of the two are equal, then:

9. The method for modeling a human upper limb projectile based on a hidden Markov model according to claim 5, characterized in that: Determine the output probability matrix, which represents the probability of the observed value corresponding to each state. The output probability matrix is ​​composed of three Gaussian distribution probability density functions: Among them, c jm Represents the coefficient of the mth Gaussian component, initialized as: c j,init ＝(0.3,0.3,0.4) For the Gaussian distribution probability density function, we have: m j,init =((1,1,1),(1,1,1),(1,1,1)) Finally, the parameter dimension determination and initialization of the hidden Markov model are completed.

10. The method for modeling a human upper limb projectile based on a hidden Markov model according to claim 9, characterized in that: Step S4 is specifically as follows: According to the forward algorithm, assuming that each forward observation is known, the probability of it belonging to state i is: α t (i)=P(o1,o2,…,o t ,s t =i|λ) Then we have: Then, by the backward algorithm, assuming that each backward observation value is known, the probability of it belonging to state i is: β t (i)=P(o t+1 ,o t+2 ,…,o T ,s t =i|λ) Then we have: Applying the Baum-Welsh algorithm, given a complete observation sequence o, the observation value o t The probability of belonging to state i is: Given a complete observation sequence o, whose state is i at time t, the probability of its state being j at time t+1 is: The final iterative formula is determined as: The iteration termination condition is: |P(o|λ i+1 )-P(o|λ i )|<0.001 The final training model is: P final =(1.0,0.0,0.0,0.0,0.0)

Images