Pengembangan Model Spasio-Temporal melalui Integrasi YOLOv11n dan Temporal Shift Module untuk Deteksi Aktivitas Kriminal Berbasis Video

Khoiriyah, Hani'atul (2026) Pengembangan Model Spasio-Temporal melalui Integrasi YOLOv11n dan Temporal Shift Module untuk Deteksi Aktivitas Kriminal Berbasis Video. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6026242003-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB)

Abstract

Pengawasan berbasis kamera CCTV memerlukan sistem deteksi aktivitas kriminal otomatis yang mampu memahami informasi spasial maupun temporal secara efisien, mengingat pendekatan spasio-temporal yang ada masih banyak dibebani komputasi berat, sementara model deteksi objek seperti YOLO masih terbatas dalam mendeteksi objek berukuran sangat kecil seperti pistol. Penelitian ini mengintegrasikan YOLOv11n dengan Temporal Shift Module (TSM) untuk pemodelan temporal, menambahkan jalur fitur P2 untuk objek kecil, serta menerapkan penyesuaian strategi pelatihan. Sebagai eksplorasi lebih lanjut, penelitian ini juga merancang varian TSM sebagai mekanisme residual bergerbang dengan parameter gate yang dapat dipelajari, sehingga besar kontribusi informasi temporal pada tiap kanal dapat ditentukan secara adaptif melalui proses pelatihan alih-alih pergeseran kanal yang tetap. Performa terbaik dicapai lewat penyesuaian strategi pelatihan dengan mAP@50 keseluruhan 0,744 pada Dataset 1 dan 0,413 pada Dataset 2. Visualisasi Motion-Aware Heatmap mengonfirmasi model memanfaatkan pola pergerakan selain karakteristik objek, dan pengujian real-time dengan integrasi Firebase serta notifikasi Telegram berhasil dilakukan. Integrasi TSM dan penambahan jalur P2 masing-masing efektif secara independen, namun kombinasi langsung keduanya justru kontraproduktif, sehingga strategi integrasi yang lebih baik antara TSM dan P2 masih diperlukan untuk penelitian selanjutnya.
==============================================================================================================================
Surveillance using CCTV cameras requires an automatic criminal activity detection system capable of efficiently understanding both spatial and temporal information, given that existing spatio-temporal approaches are still burdened by heavy computation, while object detection models such as YOLO remain limited in detecting very small objects such as handguns. This research integrates YOLOv11n with a Temporal Shift Module (TSM) for temporal modeling, adds a P2 feature path for small objects, and applies a training strategy adjustment. Results show that TSM consistently improves handgun class detection and the P2 path also improves small-object performance independently, but directly combining the two actually decreases performance. The best performance is achieved through the training strategy adjustment, with an overall mAP@50 of 0.744 on Dataset 1 and 0.413 on Dataset 2. Motion-Aware Heatmap visualization confirms that the model utilizes motion patterns in addition to object characteristics, and real-time testing with Firebase integration and Telegram notifications was successfully carried out. TSM integration and the addition of the P2 path are each independently effective, but their direct combination is counterproductive, so a better integration strategy between TSM and P2 is still needed for future research.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	YOLOv11n; Temporal Shift Module; deteksi kriminal; deteksi pistol ======================================================================================================================== YOLOv11n; Temporal Shift Module; criminal detection; handgun detection
Subjects:	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) > Information System > 59101-(S2) Master Thesis
Depositing User:	Hani`atul Khoiriyah
Date Deposited:	29 Jul 2026 06:15
Last Modified:	29 Jul 2026 06:15
URI:	http://repository.its.ac.id/id/eprint/139335

Actions (login required)

View Item