Motion capture method and apparatus, electronic device, and storage medium

1 . A motion capture method, comprising:

obtaining human inertial data collected by an extended reality device; and

obtaining human pose information by inputting the human inertial data into a pre-trained motion capture model,

wherein the motion capture model is obtained by training a neural network model based on a preset loss function, and the motion capture model comprises one or more of a neural network for predicting a human pose, a neural network for predicting global displacement information, a neural network for predicting a human bone parameter, and a classification network for predicting a human ground-stepping state, and

wherein the preset loss function is constructed based on a sliding step loss function, and the sliding step loss function is constructed by calculating a temporal motion speed for a foot that is stepping on ground.

2 . The method according to claim 1 , wherein the extended reality device comprises a head-mounted device, a handheld controller, a waist sensor, and a leg sensor,

wherein the head-mounted device, the handheld controller, the waist sensor, and the leg sensor comprise inertial sensors for collecting human inertial data.

3 . The method according to claim 1 , wherein the method further comprises, after obtaining the human inertial data collected by the extended reality device:

obtaining a feature vector for being input into the motion capture model by preprocessing the human inertial data.

4 . The method according to claim 1 , wherein training the neural network model based on the preset loss function comprises:

preprocessing and inputting human inertial data for model training into a neural network model, and fusing timing feature information of the human inertial data based on a feature fusion network of the neural network model, wherein the feature fusion network comprises a full connection layer, a normalization layer, and a recurrent neural network;

outputting a human pose parameter, global displacement information, a human bone parameter, and a human ground-stepping state through the neural network and the classification network; and

calculating a corresponding loss value based on the human pose parameter, the global displacement information, the human bone parameter, the human ground-stepping state, and the preset loss function, and adjusting a parameter of the neural network model based on the loss value until the loss value is less than a preset value, ending the training of the neural network model to obtain the motion capture model.

5 . The method according to claim 4 , wherein the preset loss function is obtained through weighted summation of a human pose prediction loss function, a root node position estimation loss function, a bone length estimation loss function, a joint point position estimation loss function, a smoothness loss function, a ground-stepping binary classification loss function, and the sliding step loss function,

wherein the human pose prediction loss function is obtained by calculating a mean squared error between a model predicted human pose parameter and a truth value pose parameter;

wherein the root node position estimation loss function is obtained by calculating a mean squared error between a model predicted root node three-dimensional coordinate parameter and a truth value root node three-dimensional coordinate parameter;

wherein the bone length estimation loss function is obtained by calculating a mean squared error between a model predicted bone parameter and a truth value bone parameter;

wherein the joint point position estimation loss function is obtained by calculating a mean squared error between a model predicted joint point position and a truth value joint point position parameter, and the model predicted joint point position is calculated by the model predicted pose parameter, the model predicted root node three-dimensional coordinate parameter, and the model predicted bone parameter; and

wherein the smoothness loss function is used to output a temporally smooth human pose result.

6 . The method according to claim 4 , wherein the human pose parameter comprise six-freedom-of-degree information of a human joint point.

7 . The method according to claim 6 , further comprising:

generating and displaying, based on the human pose parameter, a corresponding human motion in an extended reality environment.

8 . An electronic device, comprising:

at least one memory and at least one processor,

wherein the at least one memory is configured to store program code, and the at least one processor is configured to invoke the program code stored in the at least one memory to:

obtain human inertial data collected by an extended reality device; and

obtain human pose information by inputting the human inertial data into a pre-trained motion capture model,

wherein the motion capture model is obtained by training a neural network model based on a preset loss function, and the motion capture model comprises one or more of a neural network for predicting a human pose, a neural network for predicting global displacement information, a neural network for predicting a human bone parameter, and a classification network for predicting a human ground-stepping state, and

wherein the preset loss function is constructed based on a sliding step loss function, and the sliding step loss function is constructed by calculating a temporal motion speed for a foot that is stepping on ground.

9 . The electronic device according to claim 8 , wherein the extended reality device comprises a head-mounted device, a handheld controller, a waist sensor, and a leg sensor,

wherein the head-mounted device, the handheld controller, the waist sensor, and the leg sensor comprise inertial sensors for collecting human inertial data.

10 . The electronic device according to claim 8 , wherein the program code further causes the electronic device to: after obtaining the human inertial data collected by the extended reality device,

obtain a feature vector for being input into the motion capture model by preprocessing the human inertial data.

11 . The electronic device according to claim 8 , wherein the program code causing the electronic device to train the neural network model based on the preset loss function further causes the electronic device to:

preprocess and inputting human inertial data for model training into a neural network model, and fuse timing feature information of the human inertial data based on a feature fusion network of the neural network model, wherein the feature fusion network comprises a full connection layer, a normalization layer, and a recurrent neural network;

output a human pose parameter, global displacement information, a human bone parameter, and a human ground-stepping state through the neural network and the classification network; and

calculate a corresponding loss value based on the human pose parameter, the global displacement information, the human bone parameter, the human ground-stepping state, and the preset loss function, and adjusting a parameter of the neural network model based on the loss value until the loss value is less than a preset value, ending the training of the neural network model to obtain the motion capture model.

12 . The electronic device according to claim 11 , wherein the preset loss function is obtained through weighted summation of a human pose prediction loss function, a root node position estimation loss function, a bone length estimation loss function, a joint point position estimation loss function, a smoothness loss function, a ground-stepping binary classification loss function, and the sliding step loss function,

wherein the human pose prediction loss function is obtained by calculating a mean squared error between a model predicted human pose parameter and a truth value pose parameter;

wherein the root node position estimation loss function is obtained by calculating a mean squared error between a model predicted root node three-dimensional coordinate parameter and a truth value root node three-dimensional coordinate parameter;

wherein the bone length estimation loss function is obtained by calculating a mean squared error between a model predicted bone parameter and a truth value bone parameter;

wherein the joint point position estimation loss function is obtained by calculating a mean squared error between a model predicted joint point position and a truth value joint point position parameter, and the model predicted joint point position is calculated by the model predicted pose parameter, the model predicted root node three-dimensional coordinate parameter, and the model predicted bone parameter;

wherein the smoothness loss function is used to output a temporally smooth human pose result; and

wherein the ground-stepping binary classification loss function is obtained by calculating a binary cross entropy between a model predicted ground-stepping parameter and a truth value ground-stepping parameter.

13 . The electronic device according to claim 11 , wherein the human pose parameter comprise six-freedom-of-degree information of a human joint point.

14 . The electronic device according to claim 13 , the program code further causes the electronic device to:

generate and display, based on the human pose parameter, a corresponding human motion in an extended reality environment.

15 . A non-transitory computer-readable storage medium, wherein the computer-readable storage medium is configured to store program code, and the program code, when run by a computer device, causes the computer device to:

obtain human inertial data collected by an extended reality device; and

obtain human pose information by inputting the human inertial data into a pre-trained motion capture model,

wherein the motion capture model is obtained by training a neural network model based on a preset loss function, and the motion capture model comprises one or more of a neural network for predicting a human pose, a neural network for predicting global displacement information, a neural network for predicting a human bone parameter, and a classification network for predicting a human ground-stepping state, and

wherein the preset loss function is constructed based on a sliding step loss function, and the sliding step loss function is constructed by calculating a temporal motion speed for a foot that is stepping on ground.

16 . The medium according to claim 15 , wherein the extended reality device comprises a head-mounted device, a handheld controller, a waist sensor, and a leg sensor,

wherein the head-mounted device, the handheld controller, the waist sensor, and the leg sensor comprise inertial sensors for collecting human inertial data.

17 . The medium according to claim 15 , wherein the program code further causes the computer device to: after obtaining the human inertial data collected by the extended reality device,

obtain a feature vector for being input into the motion capture model by preprocessing the human inertial data.

18 . The medium according to claim 15 , wherein the program code causing the computer device to train the neural network model based on the preset loss function further causes the electronic device to:

preprocess and inputting human inertial data for model training into a neural network model, and fuse timing feature information of the human inertial data based on a feature fusion network of the neural network model, wherein the feature fusion network comprises a full connection layer, a normalization layer, and a recurrent neural network;

output a human pose parameter, global displacement information, a human bone parameter, and a human ground-stepping state through the neural network and the classification network; and

calculate a corresponding loss value based on the human pose parameter, the global displacement information, the human bone parameter, the human ground-stepping state, and the preset loss function, and adjusting a parameter of the neural network model based on the loss value until the loss value is less than a preset value, ending the training of the neural network model to obtain the motion capture model.

19 . The medium according to claim 18 , wherein the preset loss function is obtained through weighted summation of a human pose prediction loss function, a root node position estimation loss function, a bone length estimation loss function, a joint point position estimation loss function, a smoothness loss function, a ground-stepping binary classification loss function, and the sliding step loss function,

wherein the human pose prediction loss function is obtained by calculating a mean squared error between a model predicted human pose parameter and a truth value pose parameter;

wherein the root node position estimation loss function is obtained by calculating a mean squared error between a model predicted root node three-dimensional coordinate parameter and a truth value root node three-dimensional coordinate parameter;

wherein the bone length estimation loss function is obtained by calculating a mean squared error between a model predicted bone parameter and a truth value bone parameter;

wherein the joint point position estimation loss function is obtained by calculating a mean squared error between a model predicted joint point position and a truth value joint point position parameter, and the model predicted joint point position is calculated by the model predicted pose parameter, the model predicted root node three-dimensional coordinate parameter, and the model predicted bone parameter;

wherein the smoothness loss function is used to output a temporally smooth human pose result; and

wherein the ground-stepping binary classification loss function is obtained by calculating a binary cross entropy between a model predicted ground-stepping parameter and a truth value ground-stepping parameter.

20 . The medium according to claim 18 , wherein the human pose parameter comprise six-freedom-of-degree information of a human joint point.