Segmentasi Semantik pada Citra Lingkungan Lalu Lintas Menggunakan DeepLabV3+ dengan ResNet50

Azzahra, Ghefira Dhania Sarasvitha (2025) Segmentasi Semantik pada Citra Lingkungan Lalu Lintas Menggunakan DeepLabV3+ dengan ResNet50. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002211124-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Kendaraan otonom masih menghadapi tantangan dalam persepsi visual, terutama dalam kondisi ekstrem seperti pencahayaan rendah, cuaca buruk, atau silau, yang dapat menyebabkan kegagalan dalam mengenali elemen penting di lingkungan jalan. Segmentasi semantik citra menawarkan keunggulan dalam memahami konteks spasial secara menyeluruh, menjadikannya solusi potensial untuk meningkatkan ketepatan navigasi dan keselamatan kendaraan. Penelitian ini bertujuan untuk menerapkan dan menganalisis kinerja model DeepLabV3+ dengan backbone ResNet50 dalam segmentasi semantik citra lingkungan lalu lintas. Arsitektur DeepLabV3+ dengan backbone ResNet50 dipilih karena kemampuannya menangkap konteks spasial multi-skala secara efektif. Model dibangun menggunakan kombinasi dataset BDD100K sebanyak 8.000 citra dan data primer yang dikumpulkan langsung oleh penulis. Dataset diproses melalui preprocessing yang mencakup resizing, normalisasi, konversi tensor, serta diterapkan augmentasi pada data latih yang mengandung kelas minor (Horizontal Flip, Shift-ScaleRotate, Color Jitter, dan Random Brightness Contrast). Model menghasilkan segmentasi dengan 11 kelas objek lalu lintas. Penelitian ini melakukan percobaan tiga fungsi loss yang berbeda berupa Focal Loss, Dice Loss, dan Focal Tversky Loss untuk mengatasi masalah ketidakseimbangan kelas yang umum terjadi pada data segmentasi semantik. Kinerja model dievaluasi menggunakan metrik Mean Intersection over Union (mIoU), Mean Dice Coefficient, dan Pixel Accuracy. Hasil percobaan menunjukkan performa terbaik diperoleh pada model dengan Focal Loss, yang menghasilkan mIoU sebesar 0.5955, Mean Dice Coefficient sebesar 0.7056, dan Pixel Accuracy sebesar 0.9139.
=============================================================================================================================
Autonomous vehicles still face challenges in visual perception, especially under extreme conditions such as low lighting, adverse weather, or glare, which can lead to failures in recognizing important elements in road environments. Semantic image segmentation offers advantages in comprehensively understanding spatial context, making it a potential solution to improve navigation accuracy and vehicle safety. This study aims to implement and analyze the performance of the DeepLabV3+ model with a ResNet50 backbone in semantic segmentation of traffic scene images. This DeepLabV3+ architecture with a ResNet50 backbone is chosen for its ability to effectively capture multi-scale spatial context. The model is built using a combination of 8,000 images from the BDD100K dataset and primary data collected directly by the researcher. The dataset undergoes preprocessing that includes resizing, normalization, tensor conversion, and data augmentation applied to training samples containing minor classes (Horizontal Flip, Shift-Scale-Rotate, Color Jitter, and Random Brightness Contrast). The model produces segmentations with 11 traffic object classes. This study experiments with three different loss functions such as Focal Loss, Dice Loss, and Focal Tversky Loss to address class imbalance issues commonly found in semantic segmentation data. The model’s performance is evaluated using the metrics Mean Intersection over Union (mIoU), Mean Dice Coefficient, and Pixel Accuracy. The experimental results show that the best performance is achieved using Focal Loss, yielding an mIoU of 0.5955, a Mean Dice Coefficient of 0.7056, and a Pixel Accuracy of 0.9139.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Kendaraan otonom, Citra lalu lintas, Segmentasi semantik, DeepLabV3+, ResNet50, Autonomous vehicle, Traffic images, Semantic segmentation, DeepLabV3+, ResNet50
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence 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:	Ghefira Dhania Sarasvitha Azzahra
Date Deposited:	24 Jul 2025 08:53
Last Modified:	24 Jul 2025 08:53
URI:	http://repository.its.ac.id/id/eprint/121320

Actions (login required)

View Item