Fusi Heuristik Biomekanika Dan Deep Learning Untuk Rekognisi Perilaku Abnormal Multi-Person Berbasis Visi Komputer

Ariyani, Sofia (2026) Fusi Heuristik Biomekanika Dan Deep Learning Untuk Rekognisi Perilaku Abnormal Multi-Person Berbasis Visi Komputer. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 7022211014-Dissertation.pdf - Accepted Version Restricted to Repository staff only Download (29MB) | Request a copy

Abstract

Pengenalan perilaku abnormal multi-person pada lingkungan pengawasan yang padat masih menjadi tantangan yang signifikan karena oklusi yang sering terjadi dapat menimbulkan ketidakakuratan dalam proses estimasi pose manusia 3D. Sebagian besar pendekatan deep learning yang ada memproses data kerangka tubuh (skeletal) secara langsung tanpa mempertimbangkan kesesuaian dengan batasan anatomi dan biomekanika manusia, sehingga meningkatkan kemungkinan terjadinya deteksi positif palsu(false positive). Untuk mengatasi keterbatasan tersebut, penelitian ini mengusulkan suatu kerangka kerja baru berbasis BioPose, yaitu lapisan optimasi biomekanika yang mempertahankan konsistensi panjang tulang serta memastikan rentang gerak sendi tetap berada dalam batas fisiologis melalui penerapan neural inverse kinematics. Berbeda dengan pendekatan konvensional, kerangka kerja yang diusulkan terlebih dahulu memperbaiki inkonsistensi kinematika sebelum proses ekstraksi fitur dilakukan, sehingga menghasilkan representasi pose yang lebih akurat dan sesuai dengan karakteristik anatomi manusia. Selanjutnya, penelitian ini memperkenalkan mekanisme fusi fitur kinematika berbasis biomekanika yang mengintegrasikan koordinat spasial hasil koreksi dengan karakteristik gerakan temporal, meliputi kecepatan sudut sendi (joint angular velocity) dan energi kinetik (kinetic energy), menggunakan arsitektur Multi-Stream Adaptive Graph Convolutional Network (MS-AGCN). Hasil pengujian pada dataset benchmark ShanghaiTech dan CUHK Avenue menunjukkan bahwa metode yang diusulkan mampu memberikan kinerja unggul dengan mencapai nilai Area Under the Curve (AUC) masing-masing sebesar 87,6% dan 93,5%. Selain itu, model berhasil menurunkan nilai Equal Error Rate (EER) hingga 8,4%, yang menunjukkan bahwa integrasi kendala biomekanika ke dalam arsitektur deep learning secara efektif meningkatkan ketahanan model terhadap gangguan lingkungan dan oklusi, sekaligus memperbaiki akurasi pengenalan perilaku abnormal.
=====================================================================================================================================
Recognizing abnormal behavior involving multiple individuals in crowded surveillance environments remains a significant challenge because occlusions frequently introduce inaccuracies in 3D human pose estimation. Existing deep learning methods generally operate directly on raw skeletal representations, which may not satisfy anatomical and biomechanical constraints, thereby increasing the likelihood of false-positive detections. To overcome these limitations, this study presents a novel framework based on BioPose, a biomechanical optimization layer that preserves constant bone lengths and physiologically valid joint ranges through neural inverse kinematics. In contrast to conventional approaches, the proposed framework corrects kinematic inconsistencies before feature extraction, producing anatomically plausible pose representations. Furthermore, a biomechanically informed kinematic feature fusion strategy is introduced to integrate the corrected spatial joint coordinates with temporal motion descriptors, including joint angular velocity and kinetic energy, within a multi-stream adaptive graph convolutional network (MS-AGCN). Experimental evaluations on the ShanghaiTech and CUHK Avenue benchmark datasets demonstrate the effectiveness of the proposed framework, achieving AUC scores of 87.6% and 93.5%, respectively. Moreover, the model reduces the Equal Error Rate (EER) to 8.4%, indicating that incorporating biomechanical constraints into deep learning architectures substantially improves robustness against occlusion and environmental noise while enhancing abnormal behavior recognition performance.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Estimasi Pose Multi-Person, Fusi Fitur Kinematik, Heuristik Biomekanik, Rekognisi Perilaku Abnormal, Spasio-Temporal Deep Learning, Abnormal Behavior Recognition, Biomechanical Heuristics, Kinematic Feature Fusion, Pose Estimation Multi-Person, Spatio-TemporaEstimasi Pose Multi-Person, Fusi Fitur Kinematik, Heuristik Biomekanik, Rekognisi Perilaku Abnormal, Spasio-Temporal Deep Learning, Abnormal Behavior Recognition, Biomechanical Heuristics, Kinematic Feature Fusion, Pose Estimation Multi-Person, Spatio-Temporal Deep Learningl Deep Learning
Subjects:	Q Science Q Science > QA Mathematics Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20001-(S3) PhD Thesis
Depositing User:	Sofia Ariyani
Date Deposited:	05 Aug 2026 05:34
Last Modified:	05 Aug 2026 05:34
URI:	http://repository.its.ac.id/id/eprint/144045

Actions (login required)

View Item