Temporal sequence alignment method for gesture training data

1. A temporal sequence alignment method for gesture training data, comprising steps of:

step S 10 : receiving gesture training data; wherein the gesture training data comprises multiple training frames and multiple training soft labels;

step S 20 : compressing the training frames of the gesture training data to generate a compressed training frame;

step S 30 : receiving gesture sample data; wherein the gesture sample data comprises multiple sample frames and multiple sample soft labels;

step S 40 : compressing the sample frames of the gesture sample data to generate a compressed sample frame;

step S 50 : calculating an alignment model of the compressed training frame and the compressed sample frame through Dynamic Time Warping (DTW);

step S 60 : aligning the sample soft labels to multiple aligned soft labels according to the alignment model; and

step S 70 : generating an aligned training data according to the gesture sample data and the aligned soft labels.

2. The temporal sequence alignment method for the gesture training data as claimed in claim 1 , executed by a processing unit; wherein the processing unit is electrically connected to a memory unit, and the memory unit stores the gesture training data, multiple gesture sample data, and an artificial intelligence (AI) network; wherein the temporal sequence alignment method further comprises steps of:

when executing step S 10 , loading the gesture training data from the memory unit by the processing unit;

when executing step S 30 , loading one of the gesture sample data from the multiple gesture sample data stored in the memory unit by the processing unit;

after executing step S 70 , storing the aligned training data in the memory unit by the processing unit; and

when training the AI network, loading the aligned training data from the memory unit by the processing unit.

3. The temporal sequence alignment method for the gesture training data as claimed in claim 2 , wherein before step S 70 is executed, the temporal sequence alignment method further comprises the following step:

step S 65 : determining whether all of the multiple gesture sample data in the memory unit have been loaded;

when determining all of the multiple gesture sample data in the memory unit have yet been loaded, executing step S 30 to load another one of the gesture sample data from the multiple gesture sample data stored in the memory unit;

when determining all of the multiple gesture sample data in the memory unit have been loaded, executing step S 70 .

4. The temporal sequence alignment method for the gesture training data as claimed in claim 1 , wherein step S 10 further comprises the following sub-step:

step S 101 : transforming multiple training hard labels in the gesture training data corresponding to a motion picture to the training soft labels through a Cumulative Distribution Function (CDF) of Gumbel distribution.

5. The temporal sequence alignment method for the gesture training data as claimed in claim 1 , wherein step S 30 further comprises the following sub-step:

step S 301 : transforming multiple sample hard labels in the gesture sample data corresponding to a motion picture to the sample soft labels through a Cumulative Distribution Function (CDF) of Gumbel distribution.

6. The temporal sequence alignment method for the gesture training data as claimed in claim 1 , wherein:

the training frames consist of N individual training frames, and each of the individual training frames has M*M elements in a matrix;

the sample frames consist of N individual sample frames, and each of the individual sample frames has M*M elements in a matrix;

the M and N are both integers greater than one;

in each of the matrices for both the training frames and the sample frames, columns represent a speed dimension, and rows represent a distance dimension.

7. The temporal sequence alignment method for the gesture training data as claimed in claim 6 , wherein step S 20 further comprises the following sub-steps:

step S 201 : averaging speed data of the training frames to generate an averaged speed training matrix, and averaging distance data of the training frames to generate an averaged distance training matrix; and

step S 202 : combining the averaged speed training matrix and the averaged distance training matrix into the compressed training frame;

wherein the averaged speed training matrix and the averaged distance training matrix are each respectively an M*N matrix, and the compressed training frame is a 2M*N matrix; the 2M notation stands for having double an amount of M.

8. The temporal sequence alignment method for the gesture training data as claimed in claim 7 , wherein step S 40 further comprises the following sub-steps:

step S 401 : averaging speed data of the sample frames to generate an averaged speed sample matrix, and averaging distance data of the sample frames to generate an averaged distance sample matrix;

step S 402 : combining the averaged speed sample matrix and the averaged distance sample matrix into the compressed sample frame;

wherein the averaged speed sample matrix and the averaged distance sample matrix are each respectively an M*N matrix, and the compressed sample frame is a 2M*N matrix.

9. The temporal sequence alignment method for the gesture training data as claimed in claim 8 , wherein

the alignment model is an N*N matrix.

10. The temporal sequence alignment method for the gesture training data as claimed in claim 1 , wherein

the gesture training data is a training model for training to recognize a gesture; the training soft labels are time labels for recognizing a progression of the gesture.