A gesture classification method based on majority voting with adaptive voting time window length

By introducing an adaptive voting time window adjustment mechanism in most voting methods, detecting posture action category switching, the system delay problem caused by fixed window length is solved, and the accuracy and sensitivity of signal classification is improved, which is suitable for time-dependent signals. Classification tasks.

CN117725521BActive Publication Date: 2024-11-22UNIV OF ELECTRONICS SCI & TECH OF CHINA
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202311750179.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2023-09-26
Filing Date
2023-12-19
Publication Date
2024-11-22
Estimated Expiration
2043-12-19

AI Technical Summary

Technical Problem

When most existing voting methods deal with signal classification tasks with time dependencies, it is difficult to achieve high confidence and high sensitivity at the same time, especially the system delay problem caused by a fixed voting time window when switching posture actions.

Method used

Adaptive voting time window is long, and the voting time window is initialized and the switching detection time window is switched, so as to adjust the voting time window when switching posture action categories to ensure that the window length is extended when there is no switching to improve accuracy. Quick response.

Benefits of technology

It realizes the flexible adjustment of the voting time window in the posture and action classification task, reduces the system delay, improves the accuracy and response sensitivity of most voting methods, and is suitable for classification tasks such as voice signals and physiological signals.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117725521B_ABST
    Figure CN117725521B_ABST
Patent Text Reader

Abstract

The present invention provides a posture action classification method based on majority voting with an adaptive voting time window length. A classifier collects electromyographic signals to classify and identify human posture actions. The classifier outputs a classification result of a posture action every time unit and increases the length of the voting time window by one time unit. The posture action classification based on majority voting is performed within the current voting time window, and the classification result of the posture action is output. Then, it is determined whether the category switching of the posture action occurs within the switching detection time window. If so, the voting time window length is reset to an initial value and the posture action classification based on majority voting is performed again. If not, the posture action classification based on majority voting is performed directly. The present invention realizes the adaptive adjustment of the voting time window length, reduces the system delay caused by majority voting during the category switching process, and improves the accuracy of the majority voting method.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The signal classification technology of the present invention particularly relates to a classification technology based on majority voting. Background Art

[0002] In some signal classification tasks with time dependencies, such as speech recognition classification and gesture classification, the single classification result output by the classifier is often of low accuracy. The main reason is that the single classification result comes from the isolated analysis of a small segment of the signal, which does not fully utilize the time dependency of the signal, and therefore has a low confidence level.

[0003] Majority voting is a typical method for processing data stream classification. For classification tasks with time dependency, as the classification proceeds over time, a voting time window can be defined for the classification results of each time unit output by the classifier, and majority voting can be performed within the voting time window, and then the result pointed to by the majority vote is used as the final classification result. Majority voting can significantly improve the confidence of the classification results and improve the classification accuracy, and has a wide range of applications in post-classification processing algorithms.

[0004] For the classification and processing of time-dependent signals such as speech signals and biological signals, the voting time window defined by the current majority voting method is often a fixed window length. Generally speaking, the longer the window length of the voting time window, the higher the accuracy of the classification result and the higher the classification confidence. For example, by collecting electromyographic signals to classify and identify human body postures, each time unit of electromyographic signals will have an action classification. In most cases, people will not switch actions frequently. During a period of time, the human body will maintain a posture action, but there may be several time units in this period that have some other action classifications due to interference. When the human body has a continuous posture action, when the voting time window is longer, it is easy to eliminate the errors caused by interference when the classification results of most time units are correct. However, if the voting time window is fixed to a longer time, when the posture action switches, the action classification corresponding to the new action will be delayed in the final result switching due to the voting of the time units of the old action that previously accounted for the majority. Blindly extending the voting time window will lead to a decrease in system sensitivity during the category switching process. Therefore, it is difficult for the majority voting method with a fixed voting time window length to achieve high confidence and high sensitivity at the same time. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a gesture action classification method based on majority voting which takes both confidence and sensitivity into consideration by flexibly adjusting the voting time window length.

[0006] The technical solution adopted by the present invention to solve the above technical problems is a posture action classification method based on majority voting with an adaptive voting time window length, comprising the following steps:

[0007] Initialization step: Initialize the voting time window length and the switching detection time window length;

[0008] Classification step: The classifier collects electromyographic signals to classify and identify the postures and movements of the human body. The classifier outputs a classification result every time unit and increases the length of the voting time window by one time unit. The posture and movement classification is performed based on majority voting within the current voting time window, and the posture and movement classification result is output.

[0009] Switch detection time window detection step: determine whether there is a switch detection time window in the voting time window. If yes, enter the switch detection step; if not, return to the classification step; the condition for there to be a switch detection time window in the voting time window is that the current voting time window length is greater than the switch detection time window length, and the switch detection time window is a time window determined by going back one switch detection time window length from the current time unit;

[0010] Switch detection step: determine whether a gesture action category switch occurs within the switch detection time window. If so, reset the voting time window length to the initial value and return to the classification step; if not, return directly to the classification step.

[0011] The present invention realizes adaptive adjustment of the voting time window length by adding a switching detection time window to detect category switching on the basis of the traditional majority voting method with a fixed voting time window length. When the gesture action category switching is not detected, the method obtains a gesture action classification result with high accuracy and high confidence by extending the voting time window length. When the gesture action category switching is detected, the voting time window is reset to ensure that the system has high response sensitivity.

[0012] The beneficial effect of the present invention is that the voting time window length is flexibly adjusted according to the actual posture and action classification task, the system delay caused by majority voting during the posture and action category switching process is reduced, the accuracy of the majority voting method is improved, and it is suitable for signal classification tasks with time dependence such as speech signals and physiological signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a schematic diagram of adaptive adjustment of voting time window length;

[0014] Figure 2 Workflow diagram of the majority voting method for classification results with adaptive voting time window length. DETAILED DESCRIPTION

[0015] In the present invention, a switching detection time window is added to the traditional voting time window, and the switching detection time window is used to detect the switching of gesture action categories, so as to adaptively adjust the length of the voting time window. Figure 1 The overall method workflow is as shown in Figure 2 shown.

[0016] S0, initialize the voting time window length t1, switch detection time window length t2 and switch threshold Th, and take the interval time for the classifier to output a gesture action classification result as a time unit τ; t1 = 0, the setting values ​​of t2 and Th are determined according to the requirements of system sensitivity;

[0017] S1, the classifier collects electromyographic signals to classify and identify the postures and movements of the human body. The classifier outputs a posture and movement classification result every time unit τ, updates the length of the voting time window t1=t1+τ, and uses the classification result of the current time unit as one of the voting items in the voting time window;

[0018] S2. Perform majority voting based on the voting items in the voting time window, that is, each posture action classification result in the voting time window is a voting item, where there are n posture action classification results corresponding to the voting items c. k are the same, and the classification result c in the voting time window k The voting items of the gesture action classification result are the most compared with other gesture action classification results, that is, the gesture action classification result c k The most votes are obtained, and the posture action classification result c with the most votes is k Output as the final posture action classification result of the current time unit;

[0019] S3, determine whether the current voting time window length t1 is greater than the switching detection time window length t2, if so, set the [t1-t2, t1] segment window in the voting time window as the switching detection time window, and then go to step S4; if not, return to step S1;

[0020] S4. Count the posture action classification results c with the most votes in the current switching detection time window. k The number of corresponding voting items judge Is it greater than the switching threshold Th? If so, it is considered that no gesture action has switched in the current category, and the process returns to step S1; if not, it is considered that the current gesture action category has switched, and the voting time window length t1 is reset to zero, and the process returns to step S1.

Claims

1. A gesture classification method based on majority voting with adaptive voting time window length, characterized in that: The following steps are involved: Initialization step: Initialize the voting time window length t1 and the switching detection time window length t2; Classification step: the classifier collects electromyographic signals to classify and identify the postures and movements of the human body. The classifier outputs a posture and movement classification result every time unit τ and increases the length of the voting time window by one time unit t1=t1+τ; performs posture and movement classification based on majority voting within the current voting time window, and outputs the posture and movement classification result; then, determines whether the current voting time window length t1 is greater than the switching detection time window length t2, if not, returns to the classification step; If so, set the [t1-t2, t1] segment window in the voting time window as the switching detection time window; then determine whether the gesture action category switch occurs within the switching detection time window. If so, reset the voting time window length to the initial value and then perform gesture action classification based on majority voting; if not, directly perform gesture action classification based on majority voting.

2. The method according to claim 1, characterized in that: The specific method for determining the occurrence of gesture action category switching within the switching detection time window is: a switching threshold is also set in the initialization step; In the switching detection step, it is determined whether the number of the current classification results in the switching detection time window is greater than the switching threshold. If so, no gesture action category switching occurs; if not, gesture action category switching occurs.

3. The method according to claim 1, characterized in that: The initial value of the voting time window length is 0.

Citation Information

Patent Citations

  • Exoskeleton mixed brain-computer interface control method based on time coding

    CN110363157A

  • Time window length self-adaptive selection method of hybrid sight brain-computer interface

    CN112099619A

  • Human body intention recognition method based on weighted KNN

    CN113314209A

  • Method, apparatus, and system for periodic motion detection and monitoring

    EP3492945A1

  • Discovery of shifting patterns in sequence classification

    US20190303713A1