Prediksi Orientasi Tubuh Manusia Berbasis Data Point Cloud 3D Menggunakan Metode Deep Learning

Riansyah, Moch. Iskandar (2025) Prediksi Orientasi Tubuh Manusia Berbasis Data Point Cloud 3D Menggunakan Metode Deep Learning. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 7022221007_Doctoral.pdf Restricted to Repository staff only Download (10MB) | Request a copy

Abstract

Prediksi orientasi tubuh manusia berbasis data tiga dimensi merupakan komponen penting dalam sistem persepsi cerdas, seperti interaksi manusiarobot, navigasi sosial, dan pemantauan aktivitas berbasis LiDAR. Tantangan utama adalah karakteristik point cloud yang tidak terstruktur, yang menyulitkan prediksi dan mengurangi kemampuan generalisasi model deep learning. Penelitian ini mengusulkan pendekatan voxel-driven representation, yang mentransformasikan point cloud ke dalam grid 3D teratur, memungkinkan penerapan arsitektur 3D Convolutional Neural Network (3D CNN) secara efisien. Pada tahap klasifikasi, pendekatan pertama menggunakan binary voxelization, yang menunjukkan akurasi hingga 95% menggunakan VGG16 untuk empat kelas orientasi. Selanjutnya, optimasi dengan depthwise separable convolution menurunkan kompleksitas komputasi tanpa degradasi akurasi, dengan akurasi tetap 84,31% pada dataset LiDAR. Pada tahap klasifikasi lebih lanjut, pendekatan weighted voxel memperkaya representasi spasial antar voxel dengan menangkap variasi kepadatan titik, menghasilkan akurasi 98,17% pada dataset KITTI menggunakan VGG16. Kontribusi terbaru, PointVoxelNet, mengintegrasikan pembelajaran berbasis titik dengan representasi voxel melalui mekanisme residual connection dan point quantization. Evaluasi pada empat dataset ModelNet10, HumanPose,KITTI, dan TUI-NICR menunjukkan akurasi rata-rata 91% hingga 99%. Pada tahap prediksi kontinu, Voxel-BiternionNet mengatasi keterbatasan prediksi orientasi berbasis kelas dengan akurasi 6,80° MAE, mengungguli VoxNet (16,20°) dan PointNet (7,48°), serta menunjukkan ketahanan terhadap kehilangan data, dengan MAE hanya meningkat dari 6,80° menjadi 9,9° meskipun hingga 80% data hilang. Secara keseluruhan, penelitian ini menunjukkan bahwa pendekatan berbasis voxel-driven representation memberikan keseimbangan antara akurasi, robustness, dan efisiensi komputasi, menjadikannya solusi yang menjanjikan untuk sistem persepsi 3D di masa depan.
===================================================================================================================================
Human body orientation prediction based on three-dimensional data is a crucial component in intelligent perception systems, such as human-robot interaction, social navigation, and LiDAR-based activity monitoring. The main challenge lies in the unstructured nature of point clouds, which complicates prediction and reduces the generalization ability of deep learning models.This study proposes a voxel-driven representation approach, which transforms point clouds into a regular 3D grid, enabling the efficient application of 3D Convolutional Neural Network (3D CNN) architectures. In the classification phase, the first approach utilizes binary voxelization, achieving an accuracy of up to 95% using VGG16 for four orientation classes. Subsequently, optimization with depthwise separable convolution reduces computational complexity, resulting in lighter inference without sacrificing accuracy, maintaining 84.31% accuracy on the LiDAR dataset. In the next phase of classification, the weighted voxel approach enhances spatial representation between voxels by capturing variations in point density, resulting in an accuracy of 98.17% on the KITTI dataset using VGG16. The latest contribution, PointVoxelNet, integrates point-based learning with voxel representation through residual connection and point quantization mechanisms. Evaluation on four datasets ModelNet10, HumanPose, KITTI, and TUI-NICR demonstrates an average accuracy ranging from 91% to 99%. In the continuous prediction phase, Voxel-BiternionNet overcomes the limitations of class-based orientation prediction, achieving an accuracy of 6.80° MAE, outperforming VoxNet (16.20°) and PointNet (7.48°), while showing exceptional resilience to data loss, with MAE only increasing from 6.80° to 9.9°, despite up to 80% data being missing. Overall, this research demonstrates that the voxel-driven representation approach provides a balance between accuracy, robustness, and computational efficiency, making it a promising solution for future 3D perception systems.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Point Cloud, orientasi Tubuh Manusia, voxel-driven representation, Deep Learning, LiDAR, residual connection, point quantization, Point Cloud, human body orientation, voxel-driven representation, Deep Learning, LiDAR, residual connection, point quantization
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T59.7 Human-machine systems. T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7871.674 Detectors. Sensors 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:	Moch. Iskandar Riansyah
Date Deposited:	14 Jan 2026 05:32
Last Modified:	14 Jan 2026 05:32
URI:	http://repository.its.ac.id/id/eprint/129591

Actions (login required)

View Item