Deteksi Lajur untuk Mobil Otonom Pada Kondisi Lingkungan Pencahayaan Rendah dan Terdistorsi Berbasis Hybrid CNN-RNN

Airlangga, Muhammad Cendekia (2022) Deteksi Lajur untuk Mobil Otonom Pada Kondisi Lingkungan Pencahayaan Rendah dan Terdistorsi Berbasis Hybrid CNN-RNN. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 07111840000077-Undergraduate_Thesis.pdf Download (2MB)

Abstract

Deteksi lajur termasuk dalam Driving scene understanding: kemampuan mobil otonom dalam mempersepsikan lingkungan di sekitarnya tanpa bantuan manusia. Untuk melakukan deteksi ini, algoritma seperti hough transform dan convolutional neural network (CNN) telah dikembangkan. Algoritma tersebut mengalami penurunan performa yang signifikan dalam situasi tertentu seperti cuaca hujan, malam hari. Penurunan performa dari hough transform ini terjadi karena beberapa faktor: fixed region of interest (ROI) yang menyebabkan algoritma tidak bisa mendeteksi di luar ROI, hasil deteksi yang tidak adaptif, dan hasil deteksi yang terbatas hanya untuk dua garis lajur. Sementara itu, CNN memiliki kelemahan yaitu deteksi hanya dilakukan pada current frame saja. Pada penelitian ini, metode hybrid CNN-RNN digunakan untuk menyelesaikan permasalahan deteksi lajur. Kombinasi CNN-RNN ini bertujuan untuk memanfaatkan kemampuan ekstraksi fitur dari CNN dan pengolahan informasi temporal dari RNN. Dua arsitektur digunakan yaitu SegNet-ConvLSTM dan UNet-ConvLSTM. Pengujian performa dalam bentuk F1-score kedua arsitektur ini dilakukan untuk beberapa kondisi cuaca dan jalan. Secara rata-rata, UNet-ConvLSTM memiliki performa yang lebih baik. Namun, pada saat kondisi hujan, SegNet-ConvLSTM memiliki performa yang lebih baik. Hough transform menghasilkan nilai F1-score yang paling rendah. Penggunaan ConvLSTM memiliki pengaruh untuk mencegah penurunan performa yang signifikan pada saat Hujan dan Malam Hari. Pengujian terkait running time juga dilakukan pada penelitian ini. Running time dari SegNet ConvLSTM dan UNet-ConvLSTM secara berurutan adalah 0.0327 detik dan 0.0281 detik. Nilai ini menunjukkan bahwa algoritma dapat bekerja secara realtime dan jauh lebih cepat dibandingkan dengan hough transform yang memiliki running time sebesar 0.0401 detik.
===================================================================================================================================
Lane detection is one of important parts from Driving scene understanding: the ability of an autonomous car to perceive its surroundings without human assistance. Several algorithms such as the Hough transform and convolutional neural network (CNN) have been developed. The algorithm suffers from a significant decrease in performance in certain situations such as rainy weather, night. This significant performance drop in Hough transformation occurs due to several factors: fixed region of interest (ROI) which causes the algorithm to be unable to detect outside the ROI, non-adaptive detection results, and detection results that are limited to only two lanes. Meanwhile, CNN’s detection is only done on the current frame. In this study, the CNN-RNN hybrid method was used to perform a lane detection. The combination of CNN RNN aims to take advantage of the feature extraction capabilities of CNN and temporal information processing of RNN. Two architectures used are SegNet-ConvLSTM and UNet ConvLSTM. Performance testing in the form of F1-scores of these two architectures was carried out for several weather and road conditions. On average, UNet-ConvLSTM performs better. However, during rainy conditions, SegNet-ConvLSTM has better performance. Hough transformation produces the lowest F1-score value. The use of ConvLSTM has the effect of preventing a significant decrease in performance during Rain and Night. Tests related to running time were also carried out in this study. The running times of SegNet-ConvLSTM and UNet-ConvLSTM are 0.0327 seconds and 0.0281 seconds, respectively. This value indicates that the algorithm can work in real time and is much faster than the hough transform which has a running time of 0.0401 seconds.

Item Type:	Thesis (Other)
Additional Information:	RSE 006.32 Air d-1 2022
Uncontrolled Keywords:	Hybrid CNN-RNN, Mobil Otonom, Deteksi Lajur, Deep Learning, Autonomous Car, Lane Detection
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. T Technology > TL Motor vehicles. Aeronautics. Astronautics > TL152.8 Vehicles, Remotely piloted. Autonomous vehicles.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20201-(S1) Undergraduate Thesis
Depositing User:	- Davi Wah
Date Deposited:	11 Sep 2024 04:16
Last Modified:	11 Sep 2024 04:16
URI:	http://repository.its.ac.id/id/eprint/115539

Actions (login required)

View Item