چكيده به لاتين
In general, humans freely move their hands and fingers to perform functional movements or convey information. Hence, the correct recognition and classification of hand gestures has many applications such as optimal prosthetic arm design for amputees, control of upper limb prosthesis, passive rehabilitation to improve basic movements, real-time tactical sign language recognition for communicating in a safe and outdoor environment. From the visual range, it has human-computer interaction and biometric authentication in user verification. In recent years, cost reduction and increased availability of necessary hardware have made EMG signal processing a suitable solution for hand gesture recognition.
This thesis presents a new solution for recognizing and classifying hand gestures using EMG signals. This solution includes the stages of preprocessing, extraction, selection and classification of EMG data features. Since EMG signals are weak, unstable and affected by noise sources, in the pre-processing stage, a method based on the average frequency power of the Stockwell transform coefficients is used at any time to determine the area of muscle activation. Then, in order to ensure efficient real-time performance, this research divides the EMG data into 200 ms windows with 95% overlap and extracts 17 time features and 6 frequency features from each window. To calculate the frequency characteristics, a new definition is presented using the concept of power spectral density based on Stockwell transform coefficients for the time-frequency analysis of EMG signals. Also, the optimal threshold values in defining some time features are determined using AOA, GWO, SCA algorithms. The next step of the proposed solution is to normalize the features using the Z-Score method and divide them into two training and testing parts with a ratio of 70% to 30%. To determine the most accurate classifier, the training part of the data is submitted to 4 traditional machine learning algorithms, kNN, ANN, Tree and SVM, with the 5-fold cross-validation method. After identifying kNN as the superior classifier and optimizing its parameters, in order to obtain a set of features with the highest accuracy, two feature selection methods, SFS and EFS, are used, which lead to the same results. Since less muscle data of amputees can be considered as the loss of EMG receiving channel information, the final stage of the proposed solution provides a new method to determine the value of each channel's information in recognition accuracy.
In order to show the superiority of the proposed solution, experiments were conducted on the signals of 5 large databases in terms of the number of subjects and gestures, and the evaluation parameters of the confusion matrices were extracted. The minimum recognition accuracy of the optimal kNN classifier among all databases considering the best combination of temporal and frequency features was obtained as 99.84% and 99.33%, respectively. Even when the number of channels was reduced to 3, an accuracy of over 95% was achieved. These percentages indicate a significant improvement in the accuracy of the proposed method compared to the values reported in a significant number of papers in recent years. Other advantages of this method include reducing the computational load and improving the speed due to the use of the least number of features.
