Loading...

Real-time Pattern Recognition of Hand Gestures based on Machine Learning Algorithms and Surface EMG

Zandieh, Hadi | 2022

40 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55893 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Taheri, Alireza; Vossoughi, Gholamreza
  7. Abstract:
  8. The hand is an important part of the human body, the loss of all or part of it greatly reduces a person's ability to perform daily tasks. For people who have lost this significant organ, replacing it with an artificial limb that can meet some of their needs is essential. Today, all robotic prostheses use electromyographic (EMG) signals from the remaining muscles of the disabled limb as input signals. Classifying the EMG signal and converting it to a control signal faces serious challenges. Variation of signal properties over time, electrode slippage, muscle fatigue, changes in muscle contraction intensity, and changes in limb position and direction are some of these challenges. Therefore, a solution is needed that is tailored to each individual. Recently, with the increase of EMG data and the entry of this field into the big data era, the use of pattern recognition methods based on feature learning has become widespread. With recent advances, several degrees of freedom can be distinguished with high accuracy, but the performance of these models in practical applications is still unsatisfactory. One of the hopes of this field is multimode analysis, which has improved the performance of pattern recognition in this field and other fields. Therefore, 3 different datasets from the Ninapro general database including EMG signal, IMU acceleration data, and angle of claw joints have been used in this research. Data were recorded from 73 people, 13 of whom had amputations of the forearm. Different pattern recognition algorithms including three deep recurrent, convolutional and CNN-RNN networks have been investigated. The obtained results show the superiority of the accuracy and speed of the CNN neural network in identifying myoelectric patterns. The search for Optimal hyperparameters such as learning rate, batch size, and the number of nodes is performed to understand the effect of each on network performance and to select the best hyperparameters set for the final network. The final network is designed based on the results obtained from the comparison of different networks and optimal hyperparameters. The goal of the ultimate network, which is a CNN, is to detect 13 patterns of the human hand gestures with the aim of myoelectric control of the artificial limb. The accuracy of the proposed algorithm is 95.3% and the classification time of input is 3 milliseconds. EMG and IMU data with a length of 200 milliseconds were used to train the final network
  9. Keywords:
  10. Pattern Recognition ; Convolutional Neural Network ; Deep Learning ; Surface Electromyogram ; Hand ; Robotic Hand

 Digital Object List

 Bookmark

No TOC