Pengembangan Arsitektur U-Net Ringan dan Akurat untuk Segmentasi Jalan dan Kendaraan pada Autonomous Driving

Hidayat, Ferdika Pradana Putra (2025) Pengembangan Arsitektur U-Net Ringan dan Akurat untuk Segmentasi Jalan dan Kendaraan pada Autonomous Driving. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5009211026-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Perkembangan kendaraan otonom membutuhkan sistem segmentasi semantik yang akurat dan efisien secara komputasi. U-Net merupakan arsitektur populer untuk segmentasi, namun memiliki kompleksitas komputasi tinggi yang kurang ideal untuk implementasi pada perangkat terbatas. Penelitian ini mengusulkan modifikasi arsitektur U-Net menjadi versi lebih ringan dengan empat perubahan utama: (1) pengurangan channel awal dari 64 menjadi 32, (2) penggunaan Depthwise Separable Convolution, (3) penggantian transposed convolution dengan bilinear upsampling dan konvolusi 1×1, serta (4) integrasi Squeeze-and-Excitation pada skip connection. Model dievaluasi menggunakan Carla Semantic Segmentation Dataset dengan fokus pada segmentasi jalan, kendaraan, dan latar belakang. Hasil eksperimen menunjukkan U-Net ringan mempertahankan akurasi tinggi dengan mean intersection over union (mIoU) 97.93% (hanya turun 1.85% dibanding U-Net standar). U-Net ringan juga secara signifikan lebih efisien daripada U-Net standar: parameter berkurang 97% (1,04 juta berbanding 34,5 juta), ukuran model 12,5 MB (berbanding 395 MB), dan latensi 10,27 ms (4× lebih cepat). Model ini mencapai 97,36 FPS, menjadikannya cocok untuk aplikasi real-time pada sistem kendaraan otonom berbasis edge device. Penelitian ini membuktikan bahwa optimasi arsitektur U-Net dapat menghasilkan model ringan tanpa mengorbankan akurasi kritikal, dengan potensi implementasi luas di bidang otonom dan embedded vision.
=================================================================================================================================
The advancement of autonomous vehicles demands computationally efficient and accurate semantic segmentation systems. U-Net is a popular architecture for segmentation, but its high computational complexity makes it less ideal for resource-constrained devices. This study proposes a lightweight modification of the U-Net architecture with four key improvements: (1) reducing initial channels from 64 to 32, (2) employing depthwise separable convolutions, (3) replacing transposed convolution with bilinear upsampling and 1×1 convolution, and (4) integrating squeeze-and-excitation blocks into skip connections. The model was evaluated using the Carla Semantic Segmentation Dataset, focusing on segmenting roads, vehicles, and background. Experimental results demonstrate that the lightweight U-Net maintains high accuracy with a mean IoU of 97.93% (only a 1.85% drop compared to the standard U-Net) while achieving significant efficiency gains: 97% fewer parameters (1.04 million compared to 34.5 million), a model size of 12.5 MB (compared to 395 MB), and a latency of 10.27 ms (4× faster). The model achieves 97.36 FPS, making it suitable for real-time applications in autonomous driving systems on edge devices. This study proves that optimizing U-Net architecture can yield a lightweight model without sacrificing critical accuracy, with broad potential for deployment in autonomous and embedded vision systems.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	deep learning, kendaraan otonom, komputasi edge, segmentasi ; semantik, U-Net ringan, autonomous vehicles, deep learning, edge computing, lightweight U-Net, semantic segmentation.
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. R Medicine > R Medicine (General) > R858 Deep Learning T Technology > TL Motor vehicles. Aeronautics. Astronautics > TL152.8 Vehicles, Remotely piloted. Autonomous vehicles.
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Physics Engineering > 30201-(S1) Undergraduate Thesis
Depositing User:	Ferdika Pradana Putra Hidayat
Date Deposited:	31 Jul 2025 01:49
Last Modified:	31 Jul 2025 01:49
URI:	http://repository.its.ac.id/id/eprint/123271

Actions (login required)

View Item