Pengembangan Arsitektur Yolov11 Dengan Integrasi Mekanisme Attention Untuk Deteksi Truk Dengan Dimensi Berlebih (Over Dimension)

Putrawanto, Daris Irfan (2026) Pengembangan Arsitektur Yolov11 Dengan Integrasi Mekanisme Attention Untuk Deteksi Truk Dengan Dimensi Berlebih (Over Dimension). Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6026242005-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Kendaraan Over Dimension Over Load (ODOL) merupakan salah satu permasalahan dalam sistem transportasi logistik yang dapat menyebabkan kerusakan infrastruktur jalan, meningkatkan risiko kecelakaan lalu lintas, serta menurunkan efisiensi mobilitas. Proses pemantauan kendaraan ODOL masih banyak dilakukan melalui jembatan timbang maupun pengawasan manual menggunakan kamera CCTV, sehingga diperlukan sistem deteksi otomatis yang lebih efektif. Penelitian ini mengembangkan model deteksi kendaraan over dimension berbasis computer vision menggunakan arsitektur YOLOv11 yang diintegrasikan dengan tiga mekanisme attention, yaitu Shuffle Attention (SA), Mixed Local Channel Attention (MLCA), dan Efficient Local Attention (ELA). Evaluasi dilakukan menggunakan dataset Roboflow sebagai dataset utama, dataset Roboflow re-annotated, dataset CCTV Surabaya, dan dataset gabungan (mixed dataset). Kinerja model dievaluasi menggunakan metrik Precision, Recall, Mean Average Precision (mAP50 dan mAP50-95), serta waktu inferensi. Hasil eksperimen menunjukkan bahwa model usulan memberikan performa terbaik pada dataset utama dengan nilai precision sebesar 0,75, mAP50 sebesar 0,78, dan mAP50-95 sebesar 0,70, lebih tinggi dibandingkan model YOLOv11 baseline maupun model dengan mekanisme attention tunggal. Evaluasi pada dataset Roboflow re-annotated menunjukkan bahwa kualitas anotasi berpengaruh besar terhadap peningkatan performa model, sedangkan pengujian pada dataset CCTV Surabaya dan dataset gabungan menunjukkan bahwa model tetap mampu mempertahankan nilai mAP50 di atas 0,88 pada kondisi lalu lintas yang lebih kompleks. Selain itu, seluruh model memiliki waktu inferensi berkisar antara 2,7–4,2 ms per citra sehingga tetap mendukung potensi implementasi deteksi kendaraan secara real-time. Hasil penelitian menunjukkan bahwa kombinasi Shuffle Attention, MLCA, dan ELA mampu meningkatkan performa deteksi kendaraan over dimension serta memberikan representasi fitur yang lebih baik dibandingkan model baseline.
===============================================================================================================================
Over Dimension Over Load (ODOL) vehicles represent a major challenge in freight transportation systems, as they contribute to road infrastructure deterioration, increase the risk of traffic accidents, and reduce transportation efficiency. Currently, ODOL vehicle monitoring is primarily conducted through weighbridges and manual surveillance using CCTV cameras, highlighting the need for a more effective automated detection system. This study develops a computer vision-based model for over-dimension vehicle detection using the YOLOv11 architecture integrated with three attention mechanisms: Shuffle Attention (SA), Mixed Local Channel Attention (MLCA), and Efficient Local Attention (ELA). The proposed model was evaluated using the Roboflow dataset as the primary dataset, a Roboflow re-annotated dataset, the Surabaya CCTV dataset, and a mixed dataset. Model performance was assessed using Precision, Recall, Mean Average Precision (mAP50 and mAP50-95), and inference time. Experimental results show that the proposed model achieved the best performance on the primary dataset, obtaining a precision of 0.75, an mAP50 of 0.78, and an mAP50-95 of 0.70, outperforming both the YOLOv11 baseline model and models employing a single attention mechanism. Evaluation on the Roboflow re-annotated dataset demonstrates that annotation quality has a substantial impact on model performance, while experiments on the Surabaya CCTV and mixed datasets indicate that the proposed model consistently maintained an mAP50 above 0.88 under more complex traffic conditions. Furthermore, all models achieved inference times ranging from 2.7 to 4.2 ms per image, demonstrating their potential for real-time implementation. Overall, the results indicate that integrating Shuffle Attention, Mixed Local Channel Attention, and Efficient Local Attention improves over-dimension vehicle detection performance while providing more effective feature representation than the baseline model.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Kendaraan ODOL, Over Dimension, YOLOv11, Object Detection, Attention Mechanism
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 59101-(S2) Master Thesis
Depositing User:	Daris Irfan Putrawanto
Date Deposited:	29 Jul 2026 01:36
Last Modified:	29 Jul 2026 01:36
URI:	http://repository.its.ac.id/id/eprint/139166

Actions (login required)

View Item