ASHARI, Imam Ahmad and Syafei, Wahyul Amien and Wibowo, Adi (2026) INTEGRASI COMPUTER VISION DAN DEEP LEARNING UNTUK PREDIKSI REAL-TIME KEMACETAN LALU LINTAS PERKOTAAN. Doctoral thesis, UNIVERSITAS DIPONEGORO.

	Text 1.) Cover.pdf Download (443kB)
	Text 1.a) Bagian Awal.pdf Restricted to Repository staff only Download (1MB)
	Text 1.b) BAB I Pendahuluan.pdf Download (573kB)
	Text 1.c) BAB II Kajian Pustaka.pdf Download (881kB)
	Text 1.d) BAB III Metodologi Penelitian.pdf Restricted to Repository staff only Download (1MB)
	Text 1.e) BAB IV Hasil Penelitian dan Pembahasan.pdf Restricted to Repository staff only Download (1MB)
	Text 1.f) BAB V Kesimpulan dan Saran.pdf Restricted to Repository staff only Download (191kB)
	Text Daftar Pustaka.pdf Download (292kB)
	Text Lampiran.pdf Restricted to Repository staff only Download (970kB)

Abstract

Penelitian ini bertujuan untuk mengintegrasikan teknologi computer vision dan deep learning dalam sistem prediksi kemacetan lalu lintas perkotaan secara real-time. Pada tahap deteksi objek kendaraan, digunakan model YOLO dengan tiga varian yaitu YOLOv5n, YOLOv8n, dan YOLO11n. Hasil pengujian menunjukkan bahwa YOLOv8n menghasilkan performa terbaik, khususnya saat diterapkan dengan teknik multi-augmentasi (Scaling, Zoom In, Brightness Adjustment, Color Jitter, dan Noise Injection) dengan nilai mAP50–95 tertinggi sebesar 0,544. Teknik augmentasi terbukti berpengaruh signifikan terhadap peningkatan akurasi deteksi, meskipun turut berdampak pada waktu komputasi. Selanjutnya, pada tahap prediksi kemacetan, digunakan pendekatan Bi-LSTM dengan dukungan teknik sliding window dan algoritma Komodo Mlipir (KMA) untuk optimasi parameter. Model yang diusulkan, yaitu SW-KMA-Bi-LSTM, menunjukkan performa terbaik dengan nilai RMSE sebesar 14,0569 dan MAE sebesar 8,2372, mengungguli model pembanding lainnya seperti LSTM, Bi-LSTM, MFOA-Bi-LSTM, FD-Markov-LSTM, CAM-LSTM, dan Hyper-Flophet. Validasi model terbaik dilakukan pada studi kasus CCTV Sayung Demak. Hasilnya, rata-rata confidence deteksi YOLO bervariasi: 0,678 (siang) dan 0,559 (malam). Sebaliknya, prediksi SW-KMA-Bi-LSTM sangat stabil dengan rata-rata confidence 0,986 (siang) dan 0,9895 (malam). Performa YOLO (deteksi visual) dipengaruhi pencahayaan, terbukti dari penurunan confidence malam hari. SW-KMA-Bi-LSTM (prediksi data numerik) tidak terpengaruh pencahayaan, sehingga confidence tetap tinggi konsisten. Kontribusi utama penelitian ini terletak pada integrasi adaptif antara deteksi YOLOv8n dan prediksi SW-KMA-Bi-LSTM sebagai arsitektur terpadu yang dirancang untuk prediksi kemacetan real-time pada konteks lalu lintas perkotaan. Pendekatan ini tidak hanya meningkatkan akurasi, tetapi juga menghadirkan kerangka sistem yang dapat diimplementasikan langsung dalam decision support system untuk pengelolaan lalu lintas berbasis data di kota pintar.

This study aims to integrate computer vision technology and deep learning into a real-time urban traffic congestion prediction system. In the vehicle object detection stage, the YOLO model was employed with three variants: YOLOv5n, YOLOv8n, and YOLO11n. The evaluation results show that YOLOv8n achieved the best performance, particularly when applied with multi-augmentation techniques (Scaling, Zoom In, Brightness Adjustment, Color Jitter, and Noise Injection), yielding the highest mAP50–95 value of 0.544. The augmentation techniques were proven to significantly improve detection accuracy, although they also influenced computation time. Furthermore, in the congestion prediction stage, a Bidirectional LSTM approach was utilized, supported by the sliding window technique and the Komodo Mlipir Algorithm (KMA) for parameter optimization. The proposed model, SW-KMA-Bi-LSTM, demonstrated the best performance with an RMSE of 14.0569 and an MAE of 8.2372, outperforming other benchmark models such as LSTM, Bi-LSTM, MFOA-Bi-LSTM, FD-Markov-LSTM, CAM-LSTM, and Hyper-Flophet. The best-performing model was validated using a case study from the Sayung Demak CCTV footage. The results show that the average YOLO detection confidence varied between 0.678 (daytime) and 0.559 (nighttime). Conversely, the SW-KMA-Bi-LSTM predictions remained highly Stable with average confidence scores of 0.986 (daytime) and 0.9895 (nighttime). YOLO’s visual detection performance was affected by lighting conditions, as evidenced by the confidence drop during nighttime. In contrast, the SW-KMA-Bi-LSTM model, which relies on numerical data, was not influenced by lighting, resulting in consistently high confidence. The main contribution of this study lies in the adaptive integration of YOLOv8n detection and SW-KMA-Bi-LSTM prediction into a unified architecture designed for real-time congestion forecasting in urban traffic contexts. This approach not only enhances accuracy but also provides a system framework that can be directly implemented in decision support systems for data-driven traffic management in smart cities.

Actions (login required)

View Item