Klasifikasi Aktivitas Kebugaran Lansia Berdasarkan Estimasi Pose Menggunakan Deep Learning

Ulya, Amik Rafly Azmi (2024) Klasifikasi Aktivitas Kebugaran Lansia Berdasarkan Estimasi Pose Menggunakan Deep Learning. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6022221057-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2026. Download (7MB) | Request a copy

Abstract

Produktivitas orang tua sering kali menurun, terutama dalam hal kemampuan fisik. Penurunan kemampuan fisik dapat diperlambat dengan olahraga dan aktivitas fisik lainnya. Namun, aktivitas seperti peregangan sering diabaikan oleh orang tua. Aktifitas latihan untuk elderly ini menjadi penting agar elderly dapat menjaga kesehatannya di usia lanjut. Riset mengenai pengenalan aktivitas lansia telah banyak dikembangkan. Convolutional Neural Network (CNN) adalah jenis jaringan saraf tiruan yang dirancang khusus untuk pengolahan dan pengenalan gambar. Di sisi lain, Long Shoert-Term Memory (LSTM) adalah metode yang efisien untuk memecahkan masalah real-time. Kedua metode ini dapat digunakan untuk pelabelan dan pengenalan gerakan aktivitas fisik pada lansia. Aktivitas fisik yang dilakukan oleh lansia perlu disesuaikan. Dalam penelitian ini, kami mengembangkan sebuah model dengan estimasi pose lansia. Salah satu kerangka kerja untuk estimasi pose manusia adalah Mediapipe Pose Estimation (MPE). Oleh karena itu, penelitian ini berfokus pada pengenalan dan deteksi gerakan kebugaran pada lansia. Pekerjaan diawali dengan akuisisi dataset berupa aktifitas latihan yang telah disesuaikan dengan isu-isu fisik para elderly. Akusisi data dilakukan sejak sedikit dan terbatasnya dataset yang membahas aktifitas latihan ini. Data video yang telah diakuisisi kemudian dilakukan proses ekstraksi video frame. Setiap urutan frame mewakili informasi aktifitas latihan. Estimasi pose telah dilakukan menggunakan framework Mediapipe. Hasil ekstraksi ini kemudian dilatih menggunakan arsitektur CNN, LSTM, CNN-LSTM, dan deep CNN-LSTM. Akurasi setiap model sebesar 83.68%, 92.89%, 96.05%, dan 87.11%. Berdasarkan hasil tersebut, model CNN-LSTM mengungguli modelmodel lainnya dengan tingkat akurasi 96.05%. Kesalahan dalam mengenali pola data ditunjukkan menggunakan metric loss. Nilai loss model CNNLSTM sebesar 0.1498, paling kecil dibandingkan dengan model-model lainnya. Nilai ini menunjukkan kemampuan model dalam memprediksi data dengan tingkat kesalahan paling rendah. Selain itu, pada metrcis lainnya, model ini mengungguli daripada model lainnya. Precision, recall, dan f1-score model CNN-LSTM berada pada nilai 0.96, masing-masing.
=================================================================================================
Older people’s productivity often declines, especially in terms of physical ability. Physical decline can be slowed down with exercise and other physical activities. However, activities such as stretching are often neglected by the elderly. Exercise activities for the elderly are important so that the elderly can maintain their health in old age. Research on elderly activity recognition has been widely developed. Convolutional Neural Network (CNN) is a type of artificial neural network specifically designed for image processing and recognition. On the other hand, Long Time-Term Memory (LSTM) is an efficient method for solving real-time problems. These two methods can be used for labeling and recognizing physical activity movements in the elderly. Physical activities performed by the elderly need to be customized. In this study, we developed a model with elderly pose estimation. One of the frameworks for human pose estimation is Mediapipe Pose Estimation (MPE). Therefore, this research focuses on the recognition and detection of fitness movements in the elderly. The work begins with the acquisition of datasets in the form of exercise activities that have been adapted to the physical issues of the elderly. Data acquisition was carried out since there are few and limited datasets that discuss this training activity. The acquired video data was then subjected to a video frame extraction process. Each frame sequence represents the exercise activity information. Pose estimation has been done using the Mediapipe framework. The extraction results are then trained using CNN, LSTM, CNN-LSTM, and deep CNN-LSTM architectures. The accuracy of each model is 83.68%, 92.89%, 96.05%, and 87.11%. Based on these results, the CNN-LSTM model outperforms the other models with an accuracy rate of 96.05%. The error in recognizing data patterns is shown using the loss metric. The loss value of the CNN-LSTM model is 0.1498, the smallest compared to other models. This value indicates the model’s ability to predict data with the lowest error rate. In addition, in other metrics, this model outperforms other models. Precision, recall, and f1-score of the CNN-LSTM model are at 0.96, respectively.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Activity, Elderly, Pose Estimation, Aktivitas, Deep Learning, Lansia, Estimasi Pose
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Electrical Technology > Electrical Engineering > 20101-(S2) Master Thesis
Depositing User:	Amik Rafly Azmi Ulya
Date Deposited:	26 Jul 2024 07:23
Last Modified:	26 Jul 2024 07:23
URI:	http://repository.its.ac.id/id/eprint/109084

Actions (login required)

View Item