Pengembangan Model Pengenalan Aktivitas Manusia Berbasis Ekstraksi Fitur Menggunakan Deep Learning

Rahayu, Endang Sri (2024) Pengembangan Model Pengenalan Aktivitas Manusia Berbasis Ekstraksi Fitur Menggunakan Deep Learning. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 7022201010-Dissertation.pdf - Accepted Version Restricted to Repository staff only until 1 October 2026. Download (2MB) | Request a copy

Abstract

Memajukan perkembangan pengetahuan melalui penelitian pengenalan aktivitas manusia menjadi semakin penting. Deteksi yang efisien terhadap perubahan gerak suatu aktivitas akan meningkatkan akurasi pengenalan. Penelitian ini mengembangkan model baru untuk mengenali aktivitas manusia berdasarkan ekstraksi tiga jenis fitur, yaitu fitur perubahan jarak sendi, fitur pergeseran sudut sendi dan fitur percepatan gerak. Hasil ekstraksi fitur dari dataset 3D pada urutan frame di setiap segmen aktivitas dipergunakan sebagai masukan kedalam model deep learning. Ekstraksi fitur perubahan jarak sendi dikembangkan melalui perhitungan Euclidean Distance antar frame yang berdekatan pada setiap sendi. Pengembangan selanjutnya dilakukan melalui pendekatan pergeseran sudut sendi yang dikombinasikan dengan model Deep Convolutional Neural Network (DCNN). Sementara pada pengujian terhadap dataset orang lanjut usia, dibangun model yang mengintegrasikan teknik filtering menggunakan ambang batas adaptif dengan jaringan Bidirectional – Long Short-Term Memory (Bi-LSTM). Karakteristik adaptif pada ambang batas diperlukan karena setiap individu mempunyai pola aktivitas yang berbeda-beda Hasil pengujian pada ekstraksi fitur perubahan jarak sendi menunjukkan bahwa pemilihan ukuran window 16 menghasilkan akurasi model yang optimal sebesar 94,08% dalam mengklasifikasikan aktivitas manusia. Pada ekstraksi fitur pergeseran sudut sendi, kinerja model dievaluasi menggunakan matriks konfusi. Hasilnya menunjukkan bahwa model berhasil mengklasifikasikan sembilan aktivitas dalam dataset Florence 3D Actions, dengan akurasi sebesar (96,72 ± 0,83)%. Selain itu, model juga diuji pada dataset UTKinect Action3D, dan memperoleh akurasi sebesar 97,44%. Hal ini membuktikan bahwa kinerja model sangat baik telah dicapai. Sementara, pada rancangan model deep learning Bi-LSTM dengan ekstraksi fitur percepatan gerak yang menggunakan ambang batas adaptif pada pengenalan aktivitas orang lanjut usia diperoleh akurasi sebesar 94,71%. =====================================================================================================================================
Advancing knowledge development through human activity recognition rese- arch is becoming increasingly important. Efficient detection of changes in the motion of an activity will increase recognition accuracy. This research develops a new model for recognizing human activity based on extracting three types of features: joint distance change features, joint angle shift features, and movement acceleration features. The feature extraction results from the 3D dataset on the sequence of frames in each activity segment are used as input into the Deep learning model. Extraction of joint distance change features is developed through calculating Euclidean Distance between adjacent frames at each joint. Further development was carried out through a joint angle shift approach combined with a Deep Convolutional Neural Network (DCNN) model. Meanwhile, when testing a dataset of older adults, a model was built that integrated the filtering technique using Adaptive Thresholds with the Bidirectional – Long Short-Term Memory (Bi-LSTM) network. Adaptive characteristics at the threshold are needed because each individual has different activity patterns. Test results on feature extraction for changes in joint distance show that selecting a window size of 16 produces an optimal model accuracy of 94.08% in classifying human activities. The model performance is evaluated using a confusion matrix to extract joint angular displacement features. The results show that the model successfully classified nine activities in the Florence 3D Actions dataset, with an accuracy of (96.72 ± 0.83)%. In addition, the model was also tested on the UTKinect Action3D dataset and obtained an accuracy of 97.44%. So, it can be proven that excellent model performance has been achieved. Meanwhile, in the design of the Bi-LSTM Deep Learning model with motion acceleration feature extraction using adaptive thresholds for activity recognition in older adults, an accuracy of 94.71% was obtained.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Adaptive threshold, Bidirectional – Long Short-Term Memory, deep convolutional neural network, deep learning, feature extraction, human activity recognition, Ambang batas adaptif, Bidirectional – Long Short-Term Memory, deep convolutional neural network, deep learning, ekstraksi fitur, pengenalan aktivitas manusia.
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20001-(S3) PhD Thesis
Depositing User:	Endang Sri Rahayu
Date Deposited:	30 Jul 2024 04:14
Last Modified:	30 Jul 2024 04:14
URI:	http://repository.its.ac.id/id/eprint/110368

Actions (login required)

View Item