Prototype Conveyor Sederhana untuk Deteksi Objek Secara Real-time dengan Algoritma YOLOv4

Silvia Nur Anggraini, Kharis Hudaiby Hanif

Abstract


Abstract: In the digitalization era, innovation in manufacturing inventory management is essential. Conventional infrared sensor–based systems face limitations, such as a maximum detection range of 15 cm and susceptibility to lighting interference. This study proposes an automated prototype system employing the YOLOv4 deep learning algorithm for real-time object detection and counting on a production line. The system aims to automate product counting, reduce manual errors, and enhance production efficiency. Experiments were conducted on a conveyor belt at two different speeds, detecting beverage cans and snack packages. A dataset of 5,706 images was collected using a 1080P webcam positioned above the conveyor. Performance was evaluated using F-measure, accuracy, precision, and recall derived from a confusion matrix. Results indicate optimal performance at 43 RPM, achieving an average F-measure of 81% and 70.6% accuracy. Precision reached 95.7% for cans and 97.8% for snacks, with recall of 66.6% and 73.1%, respectively. At 59 RPM, the F-measure declined to 77.7% with 65.8% accuracy. YOLOv4 consistently maintained precision above 93% across all configurations, demonstrating robustness in avoiding false positives, although recall variability (62.7%–73.1%) suggests further refinement is needed to minimize false negatives. This research contributes to manufacturing automation by advancing deep learning–based inventory management systems adaptive to diverse object characteristics and operational requirements.

Abstrak: Era digitalisasi menuntut inovasi dalam pengelolaan persediaan manufaktur. Sistem konvensional berbasis sensor infrared memiliki keterbatasan dengan jangkauan deteksi maksimal 15 cm dan rentan gangguan pencahayaan. Penelitian ini mengembangkan prototype sistem otomatis menggunakan algoritma deep learning YOLOv4 untuk mendeteksi dan menghitung objek secara real-time di jalur produksi. Sistem dirancang untuk mengotomatiskan penghitungan produk, mengurangi kesalahan manual, dan meningkatkan efisiensi produksi. Pengujian dilakukan pada conveyor Belt dengan dua kecepatan berbeda untuk mendeteksi kaleng minuman dan kemasan snack. Dataset terdiri dari 5.706 citra yang diambil melalui webcam 1080P yang dipasang di atas conveyor Belt. Evaluasi menggunakan metrik F-measure, akurasi, precision, dan recall berdasarkan confusion matrix. Hasil menunjukkan performa optimal pada kecepatan 43 RPM dengan F-measure rata-rata 81% dan akurasi 70,6%. Precision mencapai 95,7% untuk kaleng minuman dan 97,8% untuk kemasan snack, dengan recall masing-masing 66,6% dan 73,1%. Pada kecepatan 59 RPM, F-measure turun menjadi 77,7% dengan akurasi 65,8%. Model YOLOv4 menunjukkan konsistensi precision tinggi di atas 93% pada semua konfigurasi, mengindikasikan reliabilitas dalam menghindari false positive. Namun, variabilitas recall 62,7%-73,1% menunjukkan perlunya perbaikan untuk meminimalkan false negative. Penelitian ini berkontribusi pada pengembangan teknologi otomatisasi manufaktur, khususnya penerapan deep learning  untuk pengelolaan inventaris yang adaptif terhadap karakteristik objek dan persyaratan operasional.


Keywords


Manufaktur; prototype; conveyor Belt; YOLOV4; deep learning

Full Text:

PDF


DOI: https://doi.org/10.35334/jbit.v5i1.6820

Refbacks

  • There are currently no refbacks.


Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

logo jbit

Address:

Gedung Dekanat, Fakultas Teknik, Universitas Borneo Tarakan. Jl. Amal Lama No. 1, Tarakan, Kalimantan Utara, Indonesia. Kodepos: 77123.
Email: jbit@borneo.ac.id

Creative Commons License

All publications

by JBIT (Jurnal Borneo Informatika dan Teknik Komputer)

are licensed under a
Creative Commons Attribution-NonCommercial 4.0 International License

External Link:
Official Web CE UBT