Implementasi Enhanced Feature Extraction Model YOLO Untuk Mendeteksi Pegawai Pada Sistem Identifikasi Pelanggaran Alat Pelindung Diri

Widyatama, Andy Kurnia (2026) Implementasi Enhanced Feature Extraction Model YOLO Untuk Mendeteksi Pegawai Pada Sistem Identifikasi Pelanggaran Alat Pelindung Diri. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 2040221052-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Pengawasan penggunaan Alat Pelindung Diri (APD) pada lingkungan industri umumnya masih dilakukan melalui pemantauan visual menggunakan kamera Closed-Circuit Television (CCTV), sehingga efektivitasnya bergantung pada kemampuan operator dalam melakukan pengawasan secara terus-menerus. Kondisi tersebut berpotensi menyebabkan keterlambatan identifikasi serta pelanggaran penggunaan APD yang tidak terdeteksi. Salah satu faktor yang mempengaruhi keberhasilan sistem identifikasi pelanggaran APD adalah kemampuan sistem dalam mendeteksi keberadaan pegawai (person) sebagai tahap awal sebelum dilakukan identifikasi atribut APD. Penelitian ini bertujuan mengembangkan sistem monitoring penggunaan APD dengan meningkatkan kemampuan person detection melalui penerapan metode Enhanced Feature Extraction menggunakan Feature Pyramid P2 (4P) dan Contrast Limited Adaptive Histogram Equalization (CLAHE) pada arsitektur YOLOv8, YOLOv11, dan YOLOv12. Evaluasi dilakukan menggunakan metrik Precision, Recall, F1-Score, mAP@0.5, mAP@0.5:0.95, Accuracy, serta Frame Per Second (FPS). Berdasarkan hasil pelatihan menggunakan test set, model YOLOv8-4P memperoleh performa terbaik dengan nilai Precision sebesar 96,24%, Recall 94,89%, F1-Score 95,56%, mAP@0.5 97,63%, dan mAP@0.5:0.95 65,29%, dengan peningkatan masing-masing sebesar 2,84%, 2,39%, 2,62%, 0,92%, dan 1,97% dibandingkan model baseline. Hasil implementasi menunjukkan bahwa model YOLOv12 memberikan performa terbaik dengan Accuracy sebesar 94,57%, Precision 98,58%, Recall 95,87%, F1-Score 97,21%, serta kecepatan inferensi 17 FPS. Pada pengujian implementasi selama 30 menit menggunakan dua sudut kamera dengan 193 pelanggaran APD berdasarkan ground truth, sistem berhasil mengidentifikasi 147 pelanggaran dengan benar, terdiri atas 136 pelanggaran rompi, 4 pelanggaran helm, dan 7 pelanggaran sepatu, serta meningkatkan jumlah pelanggaran yang terdeteksi dibandingkan sistem tanpa person detection, yaitu dari 120 menjadi 136 pelanggaran rompi, 0 menjadi 4 pelanggaran helm, dan 0 menjadi 7 pelanggaran sepatu.
====================================================================================================================================
The monitoring of Personal Protective Equipment (PPE) compliance in industrial environments is generally conducted through manual observation using Closed-Circuit Television (CCTV), making its effectiveness highly dependent on the operator's ability to continuously monitor the workplace. This condition may lead to delayed identification and undetected PPE violations. One of the key factors affecting the performance of a PPE violation identification system is its ability to detect workers (person) as the initial stage before identifying PPE attributes. This study aims to develop a PPE monitoring system by improving person detection performance through the implementation of the Enhanced Feature Extraction method using Feature Pyramid P2 (4P) and Contrast Limited Adaptive Histogram Equalization (CLAHE) on the YOLOv8, YOLOv11, and YOLOv12 architectures. Model performance was evaluated using Precision, Recall, F1-Score, mAP@0.5, mAP@0.5:0.95, Accuracy, and Frames Per Second (FPS). Based on the test set evaluation, the YOLOv8-4P model achieved the best training performance with a Precision of 96.24%, Recall of 94.89%, F1-Score of 95.56%, mAP@0.5 of 97.63%, and mAP@0.5:0.95 of 65.29%, representing improvements of 2.84%, 2.39%, 2.62%, 0.92%, and 1.97%, respectively, over the baseline model. Implementation results showed that the YOLOv12 model achieved the best performance with an Accuracy of 94.57%, Precision of 98.58%, Recall of 95.87%, F1-Score of 97.21%, and an inference speed of 17 FPS. During a 30-minute implementation test using two camera viewpoints with 193 PPE violations based on the ground truth, the proposed system correctly identified 147 violations, consisting of 136 vest violations, 4 helmet violations, and 7 safety shoe violations. Compared with the system without person detection, the number of correctly identified violations increased from 120 to 136 for safety vests, 0 to 4 for helmets, and 0 to 7 for safety shoes.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Computer Vision, Deep Learning, YOLO, Person Detection, APD, Feature Pyramid P2, CLAHE, Alat Pelindung Diri, Two-Stage Detection,Computer Vision, Deep Learning, YOLO, Person Detection, Personal Protective Equipment (PPE), Feature Pyramid P2, CLAHE, Two-Stage Detection.
Subjects:	T Technology > T Technology (General) > T55 Industrial Safety T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Vocational > 36304-Automation Electronic Engineering
Depositing User:	Andy Kurnia Widyatama
Date Deposited:	05 Aug 2026 05:40
Last Modified:	05 Aug 2026 05:40
URI:	http://repository.its.ac.id/id/eprint/144060

Actions (login required)

View Item