Sistem Peringatan Area Berbahaya Bagi Penyandang Tunanetra Berbasis Deep Learning

Widayani, Aisyah (2022) Sistem Peringatan Area Berbahaya Bagi Penyandang Tunanetra Berbasis Deep Learning. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6022201002-Master_Thesis.pdf Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Hilangnya kemampuan melihat menyebabkan seseorang sulit mengenali lingkungan sekitarnya, terutama di lingkungan yang berisiko. Orang awas dapat menghindari area yang berpotensi membahayakan dirinya dengan membaca isyarat, tetapi tidak bagi tunanetra. Penelitian ini mengusulkan sistem peringatan bagi tunanetra menggunakan teknologi visi komputer berbasis deep learning untuk mendeteksi objek. Model deep learning yang digunakan adalah YOLOv4-tiny. Sistem ini menghasilkan alarm dan notifikasi suara ketika mendeteksi tunanetra berjalan menuju lokasi berbahaya. Model deep learning mengenali citra tunanetra berdasarkan anotasi yang telah diberikan pada dataset di bagian wajah dan seluruh tubuh, dengan memegang tongkat khusus berwarna putih yang terdapat garis merah. Area berbahaya yang perlu diwaspadai oleh tunanetra yang berjalan mandiri adalah daerah bertegangan tinggi, ruang gas beracun, ruang mesin, teras di lantai tinggi, dan gerbang menuju jalan raya. Untuk mencapai tujuan ini, model YOLOv4- tiny dilatih menggunakan dataset yang terdiri dari orang normal dan tunanetra. Sistem ini dapat bekerja baik dengan pencahayaan yang cukup yaitu 7636 lux, di mana kamera dapat menangkap objek secara jelas. Metode anotasi dilakukan dengan skenario 2 kelas dan 4 kelas. Lalu, nilai average loss paling rendah dipilih untuk mendeteksi objek. Hasil terbaik dicapai skenario 4 kelas, yang terdiri dari kelas normal, normal face, blind, dan blind face, dengan average loss 0.0891. Skenario 2 kelas meliputi normal dan blind, menghasilkan average loss 0.1272. Dataset yang digunakan sebanyak 8300 citra yang terbagi menjadi 6.640 data train dan 1660 data test.
==============================================================================================================================
The loss of the ability to see causes a person to find it difficult to recognize the surrounding environment, especially in risky environments. Sighted people can avoid a potentially dangerous area by reading signs, but not the visually impaired. This study proposes a warning system for the visually impaired using deep learning- based computer vision technology to detect objects. The deep learning model used is YOLOv4-tiny. This system generates alarms and sound notifications when it detects the visually impaired walking towards a dangerous location. The deep learning model recognizes visually impaired images based on the annotations that have been given to the dataset on the face and the whole body, by holding a special white stick with a red line. Dangerous areas that blind people who walk independently need to be aware of are high voltage areas, poison gas rooms, engine rooms, terraces on high floors, and gates leading to highways. To achieve this goal, the YOLOv4-tiny model was trained using a dataset consisting of normal and visually impaired people. This system can work well with sufficient lighting, that is 7636 lux, where the camera can capture objects clearly. The annotation method is carried out with scenarios of 2 classes and 4 classes. Then, the lowest average loss value is chosen to detect objects. The best results were achieved in the 4-class scenario, which consisted of normal, normal face, blind, and blind face, classes with an average loss of 0.0891. The 2 class scenario includes normal and blind, resulting in an average loss of 0.1272. The dataset used consists of 8300 images, which are divided into 6640 training data and 1660 testing data.

Item Type:	Thesis (Masters)
Additional Information:	RTE 006.31 Wid s-1 2022
Uncontrolled Keywords:	Penyandang Tunanetra, Deep learning, Deteksi Objek, YOLOv4-tiny. Visually Impaired, Deep Learning, Object Detection, YOLOv4-tiny.
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20101-(S2) Master Thesis
Depositing User:	Mr. Marsudiyana -
Date Deposited:	06 Jul 2026 03:24
Last Modified:	06 Jul 2026 03:24
URI:	http://repository.its.ac.id/id/eprint/134331

Actions (login required)

View Item