Rekognisi Benda Rumah Tangga Menggunakan Kamera RGB-D NIR

Aiman Majid, Muhammad Ammar (2024) Rekognisi Benda Rumah Tangga Menggunakan Kamera RGB-D NIR. Other thesis, Institut Teknologi Sepuluh Nopember.

[thumbnail of 5022201197-Undergraduate_Thesis.pdf] Text
5022201197-Undergraduate_Thesis.pdf - Accepted Version
Restricted to Repository staff only until 1 October 2026.

Download (4MB) | Request a copy

Abstract

Rekognisi benda yang diaplikasikan pada robot merupakan salah satu bagian yang penting dan cukup menantang untuk dilakukan. Kesalahan rekognisi benda akan mengganggu robot dalam berinteraksi dengan benda-benda di sekitarnya. Kesalahan ini dapat disebabkan gangguan seperti pencahayaan yang kurang baik. Masalah ini dapat diatasi dengan mengambil data dari sumber lain yang juga dapat digunakan untuk mengenali benda. Pendekatan umum untuk masalah ini adalah dengan menambahkan informasi citra depth. Pada tugas akhir ini, penambahan informasi citra near infrared (NIR) dimaksudkan untuk menjadi tambahan selain citra depth saja agar rekognisi benda dapat dilakukan dengan baik meskipun dalam kondisi pencahayaan yang buruk. Tiga macam data masukan ini kemudian diproses oleh CNN masing-masing secara independen dan kemudian digabungkan dengan fusion sebelum diteruskan ke classifier untuk mengenali kelas objek. Untuk menguji kemampuan dari model, digunakan variasi pencahayaan warna putih redup, merah, hijau, biru, dan simulasi pencahayaan gelap total. Sebagai pembanding untuk model RGB-D-NIR, dibuat model RGB, depth, NIR, RGB-D, dan RGB-NIR. Dari pengujian model RGB-D-NIR, diperoleh rata-rata akurasi sebesar 89,22% dari seluruh skenario dataset pengujian.
=====================================================================================================================================
Object recognition applied to robots is one of the important and challenging parts to do. Object recognition errors will interfere with the robot in interacting with objects around it. This error can be caused by disturbances such as poor lighting. This problem can be solved by taking data from other sources that can also be used to recognize objects. A common approach to this problem is to add depth image information. In this final project, the addition of near infrared (NIR) image information is intended to be an addition to the depth image alone so that object recognition can be done well even in poor lighting conditions. The three kinds of input data are then processed by the CNN streams independently and then combined by fusion before being passed to the classifier to recognize the object class. To test the capabilities of the model, variations of dim white, red, green, blue, and total dark lighting simulations were used. As a comparison for the RGB-D-NIR model, RGB, depth, NIR, RGB-D, and RGB-NIR models were created. From testing the RGB-D-NIR model, an average accuracy of 89.22% was obtained from all test dataset scenarios.

Item Type: Thesis (Other)
Uncontrolled Keywords: Convolutional Neural Networks, Deep Learning, Kamera RGB-D, Object Recognition. Object Recognition, RGB-D Camera, Deep Learning, Convolutional Neural Networks.
Subjects: T Technology > TK Electrical engineering. Electronics Nuclear engineering
T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5105.546 Computer algorithms
T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems
Divisions: Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20201-(S1) Undergraduate Thesis
Depositing User: Muhammad Ammar Aiman Majid
Date Deposited: 30 Jul 2024 03:30
Last Modified: 30 Jul 2024 03:30
URI: http://repository.its.ac.id/id/eprint/110195

Actions (login required)

View Item View Item