Loading...

Development of a Classifier for the Human Activity Recognition System of PD Patients Using Biomechanical Features of Motion

Ejtehadi, Mehdi | 2022

211 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55033 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Behzadipour, Saeed
  7. Abstract:
  8. Parkinson’s disease (PD) is a neurodegenerative disorder and during the last few years considerable measures have been taken to rehabilitate its patients. To prevent the disorder from deteriorating and to control its progress, patients have to undergo some therapy sessions that incorporate some mobility exercises e.g. walking, sitting up and down, and etc. Since transporting the patients to the clinical centers is too burdensome, growing attention is drawn towards telerehabilitation. To this end, DMRCINT has developed a telerehab system for PD patients. This system is an intelligent classifier that uses features of linear acceleration and angular velocity signals to detect the activity that the subject is performing. These signals are captured from the user’s body segments using wearable sensors. The system can monitor the patient performing the exercises and provide the patient with real-time guidance based on his/her performance.This study aims to develop a real-time implementation of the classifier and provide a proper user interface for the use of the patients. The study also aims to investigate the effects of using biomechanical features of motion for classifying the activities. Biomechanical features are interpreted as spatio-temporal features that are more tangible and more correlated with the kinematics of human motion. Considering the new set of features, the optimal set of sensors should also be investigated.In this study, relative orientation of the body segments, and the relative orientation of a segment of the body with respect to its orientation at the standing phase were used as the biomechanical features of the motion. Similar to a previous study, moving windows of length 2.5 seconds were used in this study. Right and left thighs, and right and left forearms were proposed as the proper locations for mounting the sensors. In this study, 62 signal features were proposed to be generated for each signal. The proposed features were calculated in time domain, frequency domain, and time-frequency domain. To highlight the best generated features, Maximum Relevance Minimum Redundancy (MRMR) Algorithm was used. In order to reduce the dimensionality of the selected features space, Linear Discriminant Analysis (LDA) was used as the feature extraction algorithm. Further, Random Forest, AdaBoost, k-NN and LDA were used for the classification task. To investigate the effects of biomechanical features for classifying the activities, classifiers were evaluated when using only biomechanical features, when using only raw signal features, and when using all the features together and the results were compared.By the end of this research work, the real-time system has been implemented on the Android platform. A heuristic post-processing algorithm has been proposed that can interpret the results of the classifier and can provide proper feedback for the patient with a high accuracy. It was declared that the proposed biomechanical features that relied on the relative orientation information, could not enhance the classifier’s performance. The best classifier depended on raw signals’ features captured from the sensors mounted on left thigh and left forearm. The best classification model for the aforementioned sensors configuration and features were found to be k-NN and Random Forest (with the accuracy of 96.3%). The top 200 features for the mentioned configuration are also highlighted with their corresponding rankings
  9. Keywords:
  10. Human Activity Recognition ; Machine Learning ; Artificial Intelligence ; Wearable Sensor ; Inertial Measurement Unite (IMU) ; Parkinson Disease ; Human Movment Analysis

 Digital Object List

 Bookmark

No TOC