Loading...

Development of a Human Activity Recognition System with an Adaptive Neuro-Fuzzy Post-Processing for the Lee Silverman Voice Treatment-BIG and Functional Activities

Partovi, Ehsan | 2023

128 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55960 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Behzadipour, Saeed
  7. Abstract:
  8. Human Activity Recognition (HAR) has had tremendous improvements in the field of elderly monitoring and telerehabilitation. An anchor point for HAR systems in telerehabilitation is supervising rehabilitative excercises. For Parkinson’s disease (PD) patients, a group of rehabilitative activities, known as Lee Silverman Voice Treatment-BIG, or LSVT-BIG, have shown to be effective in improving motor performance. Similar to any rehabilitative measure, delivering these activities requires the supervision of an expert or clinician, so that the patient receives proper feedbacks. HAR systems can replace human experts. They can recognize activities and provide the user with proper feedback. HAR classifiers tend to misclassify a lot in real-time applications. The peresence of similar activities, transient activities, and parts of the activity that have not been used in the training process, are the sources of such error. This is troublesome for an HAR system that gives feedbacks to the user. Scattered, isolated misclassifications lead to consistent false alarms, which offers a bad user experience. One solution is to post-process classifier results. In this approach, the post-processor analyzed a series of classification results to determine whether the results pertain to a specific activity or not. In this study, a real-time HAR system has been developed. This system uses accelerometer and gyroscope signals from two TTGO T-Wristband smart bands, one on the left wrist, and one on the left thigh. Data are constantly sent over Bluetooth Low Energy to an Android application. 200 features are extracted from the signals. These features are then reduced to 33 features using LDA. The classifier was built by bootstrap-aggregating five LDA classifiers. All the above are handled on the Android application. The accuracy of the classifier was %94.86, but declined considerably in real-time applications. For real-time application, PPV and NPV were on average 0.44 and 0.98, respectively.To make up for the reduced performance in real-time conditions, the classifier results are post-processed. In this regard, the performance of 13 post-processing techniques has been studied. These post-processing methods include the median filter, the Hampel filter, two types of Autoencoders, Mamdani (MAMFIS) and Sugeno (ANFIS) type adaptive neuro-fuzzy inference systems, four fuzzy clustering methods, including FCM, PCM, FPCM, and PFCM, decision tree (DT), Support Vector Machine (SVM), and Isolation Forest (IF). For each activity, for five different window sizes of classifier results, and for each post-processing method, an optimization problem was defined. The goal of the optimization task was to find the best categorization of the misclassified activities as similar, transient, or irrelevant activities, with respect to a reference activity, such that PPV and NPV were maximized. The results have shown that by post-processing the classifier results using the best post-processor for each activity, for window sizes of 20, 30, 40, 50, and 60, PPV and NPV were, correspondingly, as follows: 0.88 & 0.97; 0.92 & 0.97; 0.95 & 0.98; 0.96 & 0.98; and 0.96 & 0.98. For window sizes of 20 and 30 for the classifier results, SVM had the best performance for most activities with average PPV and NPV of 0.91 & 0.97 and 0.93 & 0.98, respectively. For window sizes of 40, 50, and 60 for the classifier results, ANFIS had the best performance for most activities with average PPV and NPV of 0.95 & 0.98; 0.96 & 0.99; and 0.95 & 0.99; respectively. We conclude that post-processing can improve performance and prevent false alarms
  9. Keywords:
  10. Human Activity Recognition ; Wearable Sensor ; Parkinson Disease ; Inertial Sensor ; Machine Learning ; Neuro-Fuzzy Systems ; Post-Processing ; Rehabilitation Exercises ; Inertial Measurement Unite (IMU)

 Digital Object List

 Bookmark

No TOC