Seleksi Fitur untuk Pengenalan Gestur Tangan 3D secara Dinamis pada Interaksi Benda Virtual Menggunakan Hybrid GMM

Sooai, Adri Gabriel (2020) Seleksi Fitur untuk Pengenalan Gestur Tangan 3D secara Dinamis pada Interaksi Benda Virtual Menggunakan Hybrid GMM. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text (Feature Selection for Dynamic 3D Hand Gesture Recognition of Virtual Object Interaction Using Hybrid GMM) 07111460012001_Disertation.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Penjelajahan dan interaksi manusia dalam dunia virtual yang dilakukan berbatuan sensor 3D merambah beragam aktivitas kehidupan manusia. Khusus untuk interaksi terhadap benda virtual mengalami kemajuan pesat dalam hal ragam gestur yang digunakan. Usaha peningkatan kinerja pengenalan gestur banyak dilakukan, antara lain menggunakan ragam pengklasifikasi seperti: Naive Bayes, Quaternion Dynamic Time Warping, Hidden Conditional Neural Field, Two Stacked Long-Short Term Memory(LSTM) dan Modified LSTM melalui penambahan Reset Gate. Namun, masih terdapat celah yaitu relatif rendahnya akurasi pengenalan gestur oleh pengklasifikasi. Hal ini diakibatkan oleh tidak dimanfaatkannya seleksi fitur dalam persiapan dataset. Tujuan penelitian ini adalah menawarkan sebuah pendekatan baru untuk meningkatkan akurasi pengenalan gestur tangan melalui modifikasi seleksi fitur. Tahapan yang diusulkan dimulai dari perekaman data, menghitung simpangan baku dari tiap fitur. Kemudian, dilakukan proses pencarian korelasi positif antar fitur yang telah dihitung menggunakan fungsi probability density. Fitur berkorelasi positif dapat dilihat setelah disusun membentuk kurva gaussian mixture model(GMM) untuk kemudian digunakan dalam proses pengenalan gestur. Seluruh rangkaian proses disebut dengan HybridGMM. Pengklasifikasian menggunakan seleksi fitur Hybrid-GMM dibandingkan dengan seleksi fitur menggunakan PCA Kovarian. Diperoleh peningkatan akurasi pengenalan gestur menggunakan Hybrid-GMM pada pengklasifikasi k-NN dan SVM sebesar 99.48% dan 99.9%, mengungguli PCA Kovarian dengan selisih 0.59% dan 0.9%. Penggunaan Hybrid-GMM pada pengklasifikasi single LSTM dan tanpa penambahan Reset Gate menghasilkan akurasi 100%.

================================================================================

Exploration and interaction of humans in a virtual world that is done with a 3D sensor touch a variety of activities in human life. Especially for interactions with virtual objects experiencing rapid progress in terms of the variety of gestures used. Efforts to improve gesture recognition performance have been carried out, including using various classifiers such as Naive Bayes, Quaternion Dynamic Time Warping, Hidden Conditional Neural Fields, Two Stacked Long Short-Term Memory (LSTM) and Modified LSTM through the addition of Reset Gate. However, there is still a gap, namely the relatively low accuracy of gesture recognition by classifiers. This is caused by not using feature selection in data-set preparation. The purpose of this study is to offer a new approach to improve the accuracy of hand gesture recognition through modification of feature selection. The proposed stage starts with recording data, calculating the standard deviation of each feature. Then, the process of finding a positive correlation between features has been calculated using the probability density function. Positively correlated features can be seen after they have been arranged to form a gaussian mixture model (GMM) curve for later use in the gesture recognition process. The whole set of processes is called Hybrid-GMM. Classification using Hybrid-GMM feature selection compared to feature selection using PCA Covariance. Improved accuracy of gesture recognition using Hybrid-GMM in k-NN and SVM classifiers is 99.48% and 99.9%, ahead of PCA Covariance by 0.59% and 0.9% difference. The use of Hybrid-GMM in the single LSTM classifier and without the addition of Reset Gate produces 100% accuracy.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Hybrid-GMM, Recognition Performance Improvement, Gaussian Mixture Model, Standard Deviation, Positive Correlation, Dynamic Hand Gesture, LSTM.
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5105.546 Computer algorithms T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems
Divisions:	Faculty of Electrical Technology > Electrical Engineering > 20001-(S3) PhD Thesis
Depositing User:	Adri Gabriel Sooai
Date Deposited:	07 Sep 2020 02:15
Last Modified:	07 Sep 2020 02:15
URI:	http://repository.its.ac.id/id/eprint/81786

Actions (login required)

View Item