Deteksi Objek 3D Berbasis Monokular Dengan Variasi Pencahayaan

Husni, Alvi Zhafran (2024) Deteksi Objek 3D Berbasis Monokular Dengan Variasi Pencahayaan. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 02311940000056-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 April 2026. Download (2MB) | Request a copy

Abstract

Pendeteksian objek secara tiga dimensi atau dikenal dengan 3D object detection saat ini banyak digunakan untuk autonomous vehicle. Aspek akurasi dan real time menjadi tantangan bagi para peneliti. Untuk mendapatkan akurasi tinggi tersebut dengan memanfaatkan sensor aktif seperti LiDAR. Namun pemanfaatan LiDAR masih membutuhkan biaya yang cukup tinggi. Maka pada penelitian yang akan diusulkan, akan memanfaatkan sistem sensor pasif yaitu monokuler. Kendala sensor monokuler tidak terdapat parameter kedalaman objek atau depth. Untuk mendapatkan aspek 3D maka dilakukan rekayasa dimana akurasi yang didapatkan berdasarkan penelitian sebelumnya menggunakan sistem monokuler ini relatif rendah yaitu menghasilkan akurasi sekitar 25% untuk pendeteksian tipe hard yaitu pendeteksian dengan Intersection of Union (IoU) sebesar 0.7. Maka diusulkan solusi untuk memperbaiki isu ini yaitu dengan mengimplementasikan CNN dengan metode kedalaman diskrit dan representasi orientasi. Untuk meningkatkan performa deteksi dari susunan program dan arsitektur jaringan yang dimiliki, maka diajukan pula metode Underexposure sehingga hasil deteksi dapat meningkat.
=============================================================================================================================
Three-dimensional object detection, known as 3D object detection, is currently widely used for autonomous vehicles. Accuracy and real time aspects are a challenge for researchers. In order to get this high accuracy, many utilize active sensors such as LiDAR. However, using LiDAR requires an expensive cost. So the research that we are proposing will utilize a passive sensor system, namely monocular. The problem with monocular sensors is that there is no object depth or depth parameters. To obtain the 3D aspect, engineering was carried out where the accuracy obtained based on previous research using this monocular system was relatively low, only producing an accuracy of around 25% for a hard detection, which means an Intersection of Union (IoU) of 0.7. A solution is proposed to improve this issue, namely by implementing CNN with discrete depth methods and orientation representation. To achieve a better performance, a new method known as underexposure is implemented to increase the detection score.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	3D Object Detection, Monocular, CNN, Underexposure
Subjects:	T Technology > T Technology (General) T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Physics Engineering > 30201-(S1) Undergraduate Thesis
Depositing User:	Alvi Zhafran Husni
Date Deposited:	19 Feb 2024 06:55
Last Modified:	19 Feb 2024 06:55
URI:	http://repository.its.ac.id/id/eprint/107463

Actions (login required)

View Item