Related Experiment Videos
The application of cepstral coefficients and maximum likelihood method in EMG pattern recognition
W J Kang1, J R Shiu, C K Cheng
1Department of Electrical Engineering, National Taiwan University, R.O.C.
IEEE Transactions on Bio-Medical Engineering
|August 1, 1995
Summary
This study introduces a novel electromyographic (EMG) signal classification technique. Cepstral coefficients significantly improve movement pattern recognition compared to autoregressive coefficients, achieving over 95% accuracy with the Maximum Likelihood Method.
Area of Science:
- Biomedical Engineering
- Signal Processing
- Human Movement Analysis
Background:
- Electromyographic (EMG) signals are crucial for understanding muscle activity and movement.
- Accurate classification of EMG signals is essential for prosthetic control and rehabilitation.
- Existing feature extraction and classification methods have limitations in recognizing complex movement patterns.
Purpose of the Study:
- To develop and compare novel techniques for classifying movement patterns using EMG signals.
- To evaluate the effectiveness of cepstral coefficients versus autoregressive (AR) coefficients for feature extraction.
- To assess the performance of Euclidean Distance Measure (EDM), Weighted Distance Measure (WDM), and Maximum Likelihood Method (MLM) classifiers.
Main Methods:
- Extracted features from EMG signals using conventional autoregressive (AR) coefficients and cepstral coefficients.
- Employed three classification algorithms: EDM, WDM, and MLM, derived from the Bayes classifier.
- Modified the MLM to avoid computationally intensive matrix inversion.
- Collected EMG data from six subjects performing 10 distinct motions with surface electrodes on sternocleidomastoid and trapezius muscles.
Main Results:
- Cepstral coefficients yielded a mean recognition rate at least 5% higher than AR coefficients, a statistically significant improvement.
- The MLM classifier demonstrated the highest discrimination rate among the three algorithms.
- Combining cepstral features with MLM reduced the misclassification rate by 10.6% compared to AR coefficients with EDM.
- Selecting specific motions further increased recognition rates to over 95%.
Conclusions:
- Cepstral coefficients offer superior feature extraction for EMG-based movement classification due to better cluster separability and emphasis on informative frequency components.
- The modified MLM provides a robust and efficient classification algorithm for EMG signals.
- The proposed technique, particularly the combination of cepstral features and MLM, significantly enhances the accuracy of movement pattern recognition.