Pengembangan sistem deteksi kantuk pengemudi berbasis klasifikasi mata secara real-time menggunakan MobileNetV3

Nugroho, Septian Hadi (2026) Pengembangan sistem deteksi kantuk pengemudi berbasis klasifikasi mata secara real-time menggunakan MobileNetV3. Sarjana thesis, UIN Sunan Gunung Djati Bandung.

Preview	Text 1_cover.pdf Download (206kB) | Preview
Preview	Text 2_abstrak.pdf Download (236kB) | Preview
Preview	Text 3_skbebasplagiarism.pdf Download (569kB) | Preview
Preview	Text 4_daftarisi.pdf Download (246kB) | Preview
Preview	Text 5_bab1.pdf Download (289kB) | Preview
	Text 6_bab2.pdf Restricted to Registered users only Download (656kB) | Request a copy
	Text 7_bab3.pdf Restricted to Registered users only Download (419kB) | Request a copy
	Text 8_bab4.pdf Restricted to Registered users only Download (1MB) | Request a copy
	Text 9_bab5.pdf Restricted to Registered users only Download (236kB) | Request a copy
	Text 10_daftarpustaka.pdf Restricted to Registered users only Download (231kB) | Request a copy

Abstract

INDONESIA : Kecelakaan lalu lintas akibat kelelahan dan fenomena microsleep pengemudi masih menjadi masalah keselamatan yang kritis. Berbagai sistem pendeteksi kantuk telah dikembangkan, namun kerap menghadapi dilema antara akurasi tinggi yang membebani komputasi atau komputasi ringan namun rentan terhadap kondisi pencahayaan minim. Penelitian ini mengusulkan pengembangan sistem deteksi kantuk pengemudi secara real-time dengan mengintegrasikan pemrosesan dua tahap (two-stage pipeline). MediaPipe Face Mesh digunakan untuk melokalisasi dan mengekstraksi area mata secara cepat (<5 ms), dilanjutkan dengan klasifikasi biner status mata menggunakan arsitektur lightweight Convolutional Neural Network (CNN) MobileNetV3-Small. Model dilatih menggunakan 84.898 citra dari MRL Eye Dataset melalui pendekatan transfer learning dan on-the-fly data augmentation. Evaluasi menggunakan kerangka Design Science Research Methodology (DSRM) menunjukkan hasil yang optimal. Model klasifikasi mencapai Akurasi sebesar 96,87%, Presisi 97,65%, Recall 96,12%, F1-Score 96,88%, dan nilai AUC-ROC 0,9959. Melalui optimasi TensorFlow Lite, ukuran model berhasil dipangkas menjadi 1,96 MB dengan waktu inferensi hanya 0,71 milidetik per bingkai. Sistem diatur untuk memicu alarm peringatan kritis apabila mata terdeteksi tertutup selama 10 bingkai pemrosesan berturut-turut (ekuivalen dengan 1000 milidetik). Pengujian membuktikan sistem mampu berjalan stabil di atas 50 FPS pada spesifikasi komputasi standar dengan menerapkan frame skipping dan multithreading. Sistem ini terbukti efektif, akurat, dan efisien secara komputasi, menjadikannya sangat potensial untuk diimplementasikan sebagai fitur Advanced Driver Assistance Systems (ADAS). ENGLISH : Traffic accidents caused by driver fatigue and microsleep phenomena remain a critical safety issue. Various drowsiness detection systems have been developed, yet they frequently face a dilemma between high accuracy that burdens computation or lightweight computation that is vulnerable to low-light conditions. This study proposes the development of a real-time driver drowsiness detection system by integrating a two-stage pipeline processing. MediaPipe Face Mesh is utilized to rapidly localize and extract the eye region (<5 ms), followed by binary classification of eye status using the lightweight Convolutional Neural Network (CNN) architecture, MobileNetV3-Small. The model was trained using 84,898 images from the MRL Eye Dataset through a transfer learning approach and on-the-fly data augmentation. Evaluation using the Design Science Research Methodology (DSRM) framework showed optimal results. The classification model achieved an Accuracy of 96.87%, Precision of 97.65%, Recall of 96.12%, F1-Score of 96.88%, and an AUC-ROC value of 0.9959. Through TensorFlow Lite optimization, the model size was successfully reduced to 1.96 MB with an inference time of only 0.71 milliseconds per frame. The system is configured to trigger a critical warning alarm if the eyes are detected closed for 10 consecutive processing frames (equivalent to 1000 milliseconds). Testing proved the system capable of running stably above 50 FPS on standard computing specifications by implementing frame skipping and multithreading. This system proved to be effective, accurate, and computationally efficient, making it highly potential to be implemented as an Advanced Driver Assistance Systems (ADAS) feature.

Actions (login required)

View Item