OPTIMISASI PENDETEKSIAN OBJEK KECIL YOLOv7 UNTUK MENDETEKSI OBJEK-OBJEK AIRBORNE

Solang, Dion Andreas (2023) OPTIMISASI PENDETEKSIAN OBJEK KECIL YOLOv7 UNTUK MENDETEKSI OBJEK-OBJEK AIRBORNE. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 07211940000039-Undergraduate_Thesis.pdf Restricted to Repository staff only until 3 October 2025. Download (11MB) | Request a copy

Abstract

Pada penelitian ini, kami menunjukan percobaan kami untuk meningkatkan kemampuan deteksi YOLOv7 terhadap objek airborne. Objek-objek di airborne tampak sangat kecil pada gambar kamera ketika berada dalam jarak yang jauh. Namun, karena kecepatan pergerakan objek airborne itu tinggi, penting untuk mendeteksinya saat masih berada dalam jarak yang jauh. Oleh karena itu, YOLOv7 perlu dioptimalkan untuk dapat mendeteksi objek-objek kecil dengan baik. Dalam penelitian ini, kami mengusulkan beberapa modifikasi yang berupa perubahan dalam arsitektur (menambahkan kepala deteksi tambahan, mengalihkan skala fitur deteksi, dan mengganti head YOLO dengan decoupled anchor-free head), penerapan teknik bag-of-freebies (rekalkulasi anchor dan augmentasi mosaik), serta merubah proses inferensi (mempartisi gambar dan melakukan inferensi pada setiap partisi). Melalui eksperimen yang komprehensif, kami menemukan bahwa kombinasi penggantian head YOLO dengan decoupled anchor-free head dan melakukan inferensi pada partisi-partisi menghasilkan model yang memiliki peningkatan paling signifikan pada mean average precision (mAP) yaitu sebesar 46,18% dan tetap mempertahankan kecepatan inferensi yang real-time (> 10 FPS). Peningkatan ini jauh lebih tinggi dibandingkan dengan YOLOv7 polos tanpa modifikasi yang hanya mampu mencapai skor mAP sebesar 0%.
=============================================================================================================================
In this research, we present an attempt to improve the detection capability of YOLOv7 for airborne objects. Airborne objects appear considerably small in camera images when they are located at a considerable distance from the camera. However, due to their high speed of move ment, it is crucial to detect them while they are still far away. Therefore, to effectively detect these objects, YOLOv7 needs to be optimized for small objects. To address this challenge, we proposed several modifications that include changes in the architecture (adding an extra detection head, modifying the feature-map source, and replacing the detection head with a detached anchor-free head), application of bag-of-freebies techniques (anchor recalculation and mosaic augmentation), and change in the inference process (partitioning the image and performing inference on each partition). Through comprehensive experimentation, we have discovered that the combination of replacing the detection head with a detached anchor-free head, and performing inference on partitions yields the most promising results, with a significant increase in mean average precision (mAP) of 46.18% while still maintaining real-time inference speed (greater than 10 FPS). This improvement is notably higher compared to the unmodified plain YOLOv7, which achieved a mAP score of 0%.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deteksi Objek Kecil, YOLOv7, Modifikasi Arsitektur, Modifikasi Bag-of-Freebies, Objek Airborne Small Object Detection, YOLOv7, Architecture Modification, Bag-of-Freebies, Modification, Airborne Object
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7871.674 Detectors. Sensors T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems U Military Science > U Military Science (General) > UG Military Engineering > UG1242.D7 Unmanned aerial vehicles. Drone aircraft
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Computer Engineering > 90243-(S1) Undergraduate Thesis
Depositing User:	Dion Andreas Solang
Date Deposited:	31 Jul 2023 02:48
Last Modified:	31 Jul 2023 02:48
URI:	http://repository.its.ac.id/id/eprint/100924

Actions (login required)

View Item