Loading...

Improving the Performance of an Activity Recognition System Using Meaningful Data Augmentation and Deep Learning Methods

Riazi Bakhshayesh, Parsa | 2023

89 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55963 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Behzadipour, Saeed
  7. Abstract:
  8. Researchers working at Mowafaghian Rehabilitation Research Center have decided to develop a telerehabilitation system named SEPANTA, especially designed for activity recognition of Parkinson's Disease patients. In this regard, the system uses 34 mobility exercises, including 20 LSVT-BIG activities (especially designed for PD patients) and 14 functional daily activities. Human Activity Recognition (HAR) systems faces various challenges e.g., intra-class variabilities, meaning differences in an activity performance by different persons or a person. Data augmentation and utilizing deep learning models are the most common solutions for the risen challenges. However, deep structures require an enormous dataset. Besides, gathering experimental data from humans and patients is difficult, consumes much time and space. Consequently, data augmentation is a accessible method to encounter obstacles and make the dataset bigger and more varied. Preserving the label of the raw data (naturalness of an augmented activity) is one of the most common challenges in this field. Accordingly, researchers evaluated different augmentation methods, only mathematically comparing the changes in accuracy of a proposed model when using just raw dataset in comparison with using augmented in addition to raw data. To be mentioned, they did not compare augmented activity versus raw one, kinematically. Furthermore, they did not try to remove improper augmented data from the dataset. With this passion, the main purpose of this research is to improve the accuracy of a human activity recognition system by using meaningful data augmentation and deep learning models. To put it differently, evaluation of various augmentation methods on the performance of a deep learning model and their ability of preserving the raw label is the final goal. Additionally, the effect of the elimination of improper augmented data (using similarity measurement) should be evaluated simultaneously. To reach the goal, thesis follows two roads: using experts' opinions by watching the visualized augmented and raw activities together by a 3D model, and evaluating the effects of data augmentation and anomaly filtering of augmented data on the accuracy of the HAR system. Eventually, the results of experts' opinions indicated that the time warping is a suitable method time series augmentation, except Stand-up, Slow Walk and Fast Walk. Moreover, amplitude warping is the best choice for PO-Left and Stair Down. Also, the results of the evaluation of 21 trained model by 21 datasets (raw, filtered augmented and non-filtered augmented) represented that all possible combinations of jittering, amplitude warping and time warping performed the best among the others. Besides, using filtered datasets enhanced the accuracy of model compared to using raw dataset. The anomaly filtering significantly put positive effects on the accuracy of the HAR system other than the datasets in which time warping played a role. The accuracy improvement of models trained by filtered augmented datasets (compared with the raw dataset) were more significant on Functional activities than LSVT-BIG ones. To recognize FC-hold, the only model exhibited the accuracy of more than 95% was the model trained by the augmented datasets (filtered and not filtered) using autoencoders. In addition, to recognize FC-initial and Seated, two very similar and confusing activities, the model trained by the filtered augmented dataset using autoencoders was the only model resulted the accuracy of more than 95%
  9. Keywords:
  10. Activity Recognition ; Human Activity Recognition ; Parkinson Disease ; Wearable Sensor ; Data Augmentation ; Deep Learning ; Three-Dimensional Simulation ; Tele-rehabilitation

 Digital Object List

 Bookmark

No TOC