Jiwangga, Wiridlangit Suluh (2025) Implementasi Computer Vision untuk Deteksi Benda Tajam dan Tindakan Kekerasan Fisik di Ruang Publik Indoor Secara Real-time dengan Berbasis Edge Device. Other thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 5027211064-Undergraduate_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
5027211064-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy |
Abstract
Dalam era smart city, permintaan terhadap sistem pengawasan berbasis kecerdasan buatan terus meningkat, khususnya untuk mendeteksi aktivitas berbahaya seperti kekerasan fisik di ruang publik indoor secara real-time. Meskipun sistem konvensional seperti CCTV telah banyak diterapkan, keterbatasan dalam cakupan pengawasan serta ketergantungan pada pemantauan manual seringkali menghambat respons cepat terhadap insiden. Untuk mengatasi keterbatasan tersebut, penelitian ini mengusulkan pengembangan sistem deteksi kekerasan fisik berbasis computer vision yang dioptimalkan untuk perangkat edge computing seperti Raspberry Pi 5. Sistem ini menggunakan metode You Only Look Once (YOLO) khususnya YOLOv11 untuk mendeteksi keberadaan objek tajam dan perilaku kekerasan, serta dirancang agar tetap efisien pada perangkat dengan daya komputasi rendah tanpa mengorbankan akurasi deteksi di berbagai kondisi lingkungan. Selain itu, sistem ini terintegrasi dengan Firebase untuk penyimpanan data secara real-time dan dilengkapi dengan dashboard interaktif berbasis Streamlit yang memungkinkan pengguna meninjau bukti visual, metadata, dan log deteksi secara langsung. Model terbaik setelah dilakukan tuning hyperparameter berhasil mencapai nilai mean Average Precision (mAP) sebesar 91,7% pada IoU 0,50 dan 65,2% pada mAP50-95, dengan precision sebesar 88,2% dan recall sebesar 87,2%. Jika dibandingkan dengan model baseline yang menghasilkan mAP50 sebesar 90,8% dan mAP50-95 sebesar 63,6%, model yang telah dituning menunjukkan peningkatan performa yang signifikan. Hasil ini membuktikan bahwa sistem yang dikembangkan mampu mendeteksi kekerasan fisik dan keberadaan senjata tajam secara akurat dan efisien dalam lingkungan indoor secara real-time. Meskipun demikian, faktor seperti kondisi pencahayaan dan efek blur dapat memengaruhi performa deteksi.
========================================================================================================================================
In the era of smart cities, the demand for artificial intelligence-based surveillance systems is growing, particularly for detecting hazardous activities such as physical violence in indoor public spaces in real time. Although conventional systems like CCTV are widely
implemented, their limited coverage and dependence on manual observation often hinder prompt responses to incidents. To overcome these limitations, this study proposes the
development of a computer vision-based physical violence detection system, optimized for edge computing devices such as the Raspberry Pi 5. The system employs the You Only Look
Once (YOLO) especially YOLOv11 method to detect both sharp objects and violent behavior, and is designed to operate efficiently on low-power hardware while maintaining detection accuracy across various environmental conditions. In addition, the system integrates with Firebase for real-time data storage and features an interactive monitoring dashboard developed using Streamlit, enabling users to review visual evidence, metadata, and detection
logs. The best-performing model, after hyperparameter tuning, achieved a mean Average Precision (mAP) of 91.7% at IoU 0.50 and 65.2% at mAP50-95, with a precision of 88.2%
and a recall of 87.2%. Compared to the baseline model, which yielded mAP50 of 90.8% and mAP50-95 of 63.6%, the tuned model showed a measurable improvement. These results
demonstrate that the proposed system can detect both physical violence and the presence of sharp objects accurately and efficiently in real-time indoor environments. Nonetheless, factors such as lighting conditions and image blur can influence detection performance
Actions (login required)
 |
View Item |