Deteksi Jatuh Berbasis Skeleton Menggunakan Pose Estimation YOLO11 dan Block Graph Convolutional Network

Juan, Adnan Abdullah (2026) Deteksi Jatuh Berbasis Skeleton Menggunakan Pose Estimation YOLO11 dan Block Graph Convolutional Network. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221155-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (10MB) | Request a copy

Abstract

Jatuh merupakan permasalahan kesehatan serius yang dapat menyebabkan cedera fisik serta komplikasi fisiologis seperti long lie dan post-fall syndrome. Deteksi dini kejadian jatuh diperlukan untuk meningkatkan keselamatan pengguna. Pendekatan skeleton-based merepresentasikan tubuh manusia sebagai sekumpulan titik kunci hasil pose estimation yang digunakan sebagai dasar analisis aktivitas. Metode Spatial-Temporal Graph Convolutional Network (ST-GCN) telah banyak digunakan untuk pengenalan aktivitas berbasis skeleton, namun masih bergantung pada topologi graf statis. Penelitian ini mengembangkan sistem deteksi jatuh yang mengintegrasikan YOLO11-pose untuk ekstraksi 17 keypoint tubuh dan Block Graph Convolutional Network (BlockGCN) untuk klasifikasi biner jatuh dan tidak jatuh. Evaluasi dilakukan menggunakan dataset NTU RGB pada empat konfigurasi, yaitu kombinasi representasi 17 keypoint COCO dan 25 keypoint NTU RGB dengan komposisi balanced dan imbalanced. Karena data uji tidak seimbang, kinerja diukur menggunakan balanced accuracy sebagai metrik utama. Konfigurasi terbaik diperoleh pada 17 keypoint dengan komposisi imbalanced menggunakan weighted cross-entropy loss yang mencapai accuracy 99,20%, precision 99,08%, recall 98,62%, dan F1-score 98,85%. Model tergolong ringan dengan 1.266.562 parameter serta 1,198 GFLOPs dan mampu beroperasi near real-time pada sekitar 28,4 FPS dengan latensi 22,71 ms per frame dan memori GPU 120,5 MB, sehingga akurat sekaligus efisien untuk perangkat komputasi umum.
===================================================================================================================================
Falls are a serious health problem that can cause physical injuries as well as physiological complications such as long lie and post-fall syndrome. Early detection of fall events is needed to improve user safety. The skeleton-based approach represents the human body as a set of keypoints obtained from pose estimation, which serve as the basis for activity analysis. The Spatial-Temporal Graph Convolutional Network (ST-GCN) has been widely used for skeleton-based action recognition but still relies on a static graph topology. This study develops a fall detection system that integrates YOLO11-pose for extracting 17 body keypoints and a Block Graph Convolutional Network (BlockGCN) for the binary classification of fall and non-fall. The evaluation was conducted on the NTU RGB dataset across four configurations, namely the combination of 17-keypoint COCO and 25-keypoint NTU RGB representations with balanced and imbalanced compositions. Since the test data is imbalanced, performance was measured using balanced accuracy as the primary metric. The best configuration was obtained with 17 keypoints and the imbalanced composition trained using weighted cross-entropy loss, achieving an accuracy of 99.20%, precision of 99.08%, recall of 98.62%, and F1-score of 98.85%. The model is lightweight, with 1,266,562 parameters and 1.198 GFLOPs, and is capable of operating in near real-time at about 28.4 FPS with a latency of 22.71 ms per frame and 120.5 MB of GPU memory, making it both accurate and efficient for general purpose computing devices.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deteksi Jatuh, YOLO11, Block Graph Convolutional Network, Representasi skeleton, Near real-time, Fall detection, YOLO11, Block Graph Convolutional Network, Skeleton representation, Near real-time
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Adnan Abdullah Juan
Date Deposited:	24 Jul 2026 02:03
Last Modified:	24 Jul 2026 02:03
URI:	http://repository.its.ac.id/id/eprint/136997

Actions (login required)

View Item