PERBANDINGAN FASTER R-CNN DAN YOLOv8 UNTUK DETEKSI API DAN ASAP PADA DATA VIDEO

Gumara, Aditya Dwi (2026) PERBANDINGAN FASTER R-CNN DAN YOLOv8 UNTUK DETEKSI API DAN ASAP PADA DATA VIDEO. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002201070_Undergraduate_Thesis.pdf Download (5MB)

Abstract

Kebakaran merupakan salah satu bencana yang dapat menimbulkan kerugian besar dari segi ekonomi, lingkungan, maupun keselamatan manusia. Beberapa peristiwa kebakaran besar, seperti kebakaran Gedung Kejaksaan Agung pada tahun 2020 dan kebakaran kilang minyak Pertamina Cilacap pada tahun 2021, menunjukkan pentingnya sistem deteksi dini yang mampu mendeteksi kebakaran secara cepat dan akurat. Sistem deteksi kebakaran konvensional yang memanfaatkan sensor api, gas, atau suhu masih memiliki keterbatasan, seperti cakupan area yang sempit dan keterlambatan dalam mendeteksi sumber kebakaran. Oleh karena itu, penelitian ini mengusulkan pendekatan berbasis computer vision menggunakan metode deep learning untuk mendeteksi objek api (fire) dan asap (smoke) pada data video. Model yang digunakan adalah Faster R-CNN ResNet-50 FPN dan YOLOv8m. Tahapan penelitian meliputi pengumpulan dan anotasi data video, pelatihan model melalui proses hyperparameter tuning, serta pengujian model pada berbagai skenario kondisi pencahayaan. Hasil hyperparameter tuning menunjukkan bahwa konfigurasi terbaik Faster R-CNN ResNet-50 FPN diperoleh menggunakan optimizer SGD dengan learning rate 0,1, menghasilkan mAP50 sebesar 75,57% dan mAP50-95 sebesar 43,10%. Sementara itu, konfigurasi terbaik YOLOv8m diperoleh menggunakan optimizer AdamW dengan learning rate 0,001, menghasilkan mAP50 sebesar 77,95% dan mAP50-95 sebesar 44,94%. Pada tahap pengujian, YOLOv8m secara umum menghasilkan nilai Precision, Recall, dan F1-Score yang lebih tinggi dibandingkan Faster R-CNN ResNet-50 FPN serta menunjukkan performa yang lebih stabil pada berbagai kondisi pencahayaan. Selain itu, YOLOv8m memiliki kecepatan inferensi sekitar 41–43 FPS, sedangkan Faster R-CNN ResNet-50 FPN hanya mencapai sekitar 10–11 FPS. Berdasarkan hasil tersebut, dapat disimpulkan bahwa YOLOv8m merupakan model yang lebih sesuai untuk diterapkan pada sistem deteksi dini kebakaran berbasis video secara real-time karena memberikan keseimbangan yang lebih baik antara akurasi deteksi dan kecepatan inferensi.
===================================================================================================================================
Fires are one of the disasters that can cause significant losses in terms of the economy, the environment, and human safety. Several major fire incidents, such as the fire at the Attorney General’s Office building in 2020 and the fire at the Pertamina Cilacap oil refinery in 2021, highlight the importance of an early-warning system capable of detecting fires quickly and accurately. Conventional fire detection systems that rely on flame, gas, or temperature sensors still have limitations, such as limited coverage and delays in detecting the source of a fire. Therefore, this study proposes a computer vision-based approach using deep learning methods to detect fire and smoke objects in video data. The models used are Faster R-CNN ResNet-50 FPN and YOLOv8m. The research stages include video data collection and annotation, model training through hyperparameter tuning, and model testing under various lighting conditions. The results of the hyperparameter tuning show that the best configuration for Faster R-CNN ResNet-50 FPN was obtained using the SGD optimizer with a learning rate of 0.1, yielding an mAP50 of 75.57% and an mAP50-95 of 43.10%. Meanwhile, the best configuration for YOLOv8m was obtained using the AdamW optimizer with a learning rate of 0.001, yielding an mAP50 of 77.95% and an mAP50-95 of 44.94%. Meanwhile, the best configuration for YOLOv8m was obtained using the AdamW optimizer with a learning rate of 0.001, yielding an mAP50 of 77.95% and an mAP50-95 of 44.94%. During the testing phase, YOLOv8m generally produced higher Precision, Recall, and F1-Score values compared to Faster R-CNN ResNet-50 FPN and demonstrated more stable performance under various lighting conditions. Furthermore, YOLOv8m has an inference speed of approximately 41–43 FPS, whereas Faster R-CNN ResNet-50 FPN only reaches about 10–11 FPS. Based on these results, it can be concluded that YOLOv8m is a more suitable model for implementation in real-time, video-based early fire detection systems because it provides a better balance between detection accuracy and inference speed.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	kebakaran, deteksi api dan asap, deep learning, Faster R-CNN, YOLOv8, video fire, fire and smoke detection, deep learning, Faster R-CNN, YOLOv8, video
Subjects:	Q Science Q Science > QA Mathematics Q Science > QA Mathematics > QA76.6 Computer programming.
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Aditya Dwi Gumara
Date Deposited:	04 Aug 2026 07:28
Last Modified:	04 Aug 2026 07:28
URI:	http://repository.its.ac.id/id/eprint/143533

Actions (login required)

View Item