Deteksi Nyala Lampu Lalu Lintas Secara Real-Time Menggunakan Tiny-YOLOv7

Pradifto, Bagasetya (2025) Deteksi Nyala Lampu Lalu Lintas Secara Real-Time Menggunakan Tiny-YOLOv7. Other thesis, InstitutTeknologi Sepuluh Nopember.

Text 5002211087-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (24MB) | Request a copy

Abstract

Seiring dengan perkembangan teknologi, inovasi di bidang transportasi terus bermunculan, salah satunya adalah kendaraan otonom yang dirancang untuk beroperasi tanpa intervensi manusia. Agar dapat berfungsi secara aman dan efisien, kendaraan otonom membutuhkan sistem deteksi yang andal untuk mengenali elemen-elemen penting di jalan, termasuk status nyala lampu lalu lintas secara real-time. Deteksi yang tidak akurat dapat mengganggu proses pengambilan keputusan dan berpotensi menimbulkan risiko keselamatan. Untuk mengatasi tantangan tersebut, penelitian ini mengimplementasikan algoritma Tiny-YOLOv7, sebuah varian YOLO yang ringan namun mampu memberikan keseimbangan antara kecepatan dan akurasi. Penelitian ini bertujuan untuk melakukan evaluasi performa Tiny-YOLOv7 dalam mendeteksi nyala lampu lalu lintas berdasarkan tiga kelas warna (merah, kuning, hijau) di berbagai kondisi cuaca. Proses penelitian mencakup tahap pengumpulan dataset melalui perekaman langsung di wilayah Surabaya dengan berbagai kondisi cuaca, preprocessing data mencakup anotasi data menggunakan format YOLO dan augmentasi data untuk memperkaya variasi data, pelatihan model Tiny-YOLOv7 dengan menggunakan backbone ELAN-S dan neck PANet simplified, Model dievaluasi berdasarkan metrik precision, recall, mAP50, mAP50:95, serta kecepatan pemrosesan (FPS). Hasil penelitian menunjukkan bahwa model Tiny-YOLOv7 mencapai akurasi deteksi yang baik ditunjukan dengan nilai mAP50 sebesar 99,6% dan mAP50:95 sebesar 62,0% pada kondisi cerah. Performa deteksi model hanya sedikit menurun pada kondisi berawan, dengan nilai mAP50 sebesar 99.1% dan mAP50:95 sebesar 44.7%, namun performa deteksinya menurun drastis pada kondisi hujan dengan nilai mAP 34,2% dan mAP50:95 sebesar 10,6%. Model ini juga mampu melakukan enghasilkan deteksi dengan rata-rata kecepatan inferensi >300 FPS (2,1–2,8 ms per frame), hal ini menunjukkan bahwa model memiliki potensi besar untuk penerapan deteksi lampu lalu lintas real-time.
====================================================================================================================================
With the advancement of technology, innovations in transportation continue to emerge, one of which is autonomous vehicles designed to operate without human intervention. In order to function safely and efficiently, autonomous vehicles require a reliable detection system to recognize important elements on the road, including the status of traffic lights in real time. Inaccurate detection can disrupt decision-making processes and potentially pose safety risks. To address this challenge, this study implements the Tiny-YOLOv7 algorithm, a lightweight variant of YOLO that balances speed and accuracy. This study aims to evaluate the performance of Tiny-YOLOv7 in detecting traffic light status based on three color classes (red, yellow, green) under various weather conditions. The research process includes the collection of datasets through direct recording in the Surabaya area under various weather conditions, pre-processing of data including data annotation using the YOLO format and data augmentation to enrich data variation, training of the Tiny-YOLOv7 model using the ELAN-S backbone and simplified PANet neck, and evaluation of the model based on precision, recall, mAP50, mAP50:95, and processing speed (FPS). The research results show that the Tiny-YOLOv7 model achieves good detection accuracy, as indicated by an mAP50 value of 99.6% and an mAP50:95 value of 62.0% under clear conditions. The model’s detection performance only slightly decreases under cloudy conditions, with an mAP50 value of 99.1% and an mAP50:95 value of 44.7%, but its detection performance drops significantly under rainy conditions, with an mAP50 value of 34.2% and an mAP50:95 value of 10.6%. This model is also capable of generating detections with an average inference speed of 300 FPS (2.1–2.8 ms per frame), indicating that the model has great potential for real-time traffic light detection applications.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Tiny-YOLOv7, YOLO, Kendaraan Otonom, Deteksi, Real-Time, Tiny-YOLOv7, YOLO, Autonomous Vehicles, Detection, Real-Time
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.6 Computer programming. Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Bagasetya Dwiki Pradifto
Date Deposited:	31 Jul 2025 03:22
Last Modified:	31 Jul 2025 03:22
URI:	http://repository.its.ac.id/id/eprint/123408

Actions (login required)

View Item