Analisis Algoritma SSD MobileNet V2 dan YOLOv4-tiny untuk Human Detection pada Amlogic S905x

Narain, Gavin Bagus Kenzie (2023) Analisis Algoritma SSD MobileNet V2 dan YOLOv4-tiny untuk Human Detection pada Amlogic S905x. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 05311940000028-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 September 2025. Download (4MB) | Request a copy

Abstract

Edge computing memungkinkan proses komputasi diletakkan sedekat mungkin ke lokasi di mana komputasi tersebut dibutuhkan. Dengan berkembangnya sistem tertanam yang terkoneksi internet saat ini, edge computing memiliki keunggulan dalam latensi dan volume transfer data. Aplikasi human detection, memiliki jumlah transfer data video yang banyak, tentunya dapat diterapkan ke edge computing untuk menghemat sumber daya. Pada penelitian ini, akan dilakukan implementasi human detection pada salah satu perangkat dengan sumber daya terbatas, yaitu Amlogic S905x. Implementasi akan dilakukan dengan dua algoritma object detection yang populer karena performanya untuk mobile device, yaitu SSD MobileNet V2 dan YOLOv4-tiny. Implementasi akan dibagi menjadi 2 bagian, melalui rekaman video dan secara real-time melalui RTSP. Kedua model di-train pada dataset MS COCO 2017 dengan pengecualian hanya mengambil kelas person. Hasil training model didapatkan nilai AP@0.50 sebesar 0.596 untuk SSD MobileNet V2 dan AP@0.50 sebesar 0.5268 untuk YOLOv4-tiny. Kemudian, pada implementasi yang dilakukan pada rekaman video, SSD MobileNet V2 dan YOLOv4-tiny mendapat rerata FPS sebesar 1.98 dan 3.49 secara berturutan. Untuk implementasi secara real-time dilakukan dengan bantuan website untuk menampilkan gambar deteksi dari kamera IP secara real-time. Pada implementasi real-time ini, SSD MobileNetV2 dan YOLOv4-tiny mendapat rerata FPS sebesar 1.92 dan 3.47 secara berturutan.
=============================================================================================================================
Edge computing allows computing processes to be placed as close as possible to the location where they are needed. With the development of internet-connected embedded systems today, edge computing has advantages in latency and data transfer volume. Human detection applications, which have a large amount of video data transfer, can certainly be applied to edge computing to save resources. In this research, we will implement human detection on one of the devices with limited resources, the Amlogic S905x. The implementation will be done with two object detection algorithms that are popular for their performance for mobile devices, namely MobileNet V2 SSD and YOLOv4-tiny. The implementation will be divided into 2 parts, through video recording and in real-time through RTSP. Both models are trained on the MS COCO 2017 dataset with the exception of only taking the person class. The results of the model training obtained a value AP@0.50 of 0.596 for SSD MobileNet V2 and AP@0.50 of 0.5268 for YOLOv4-tiny. Then, in the implementation carried out on video recordings, SSD MobileNet V2 and YOLOv4-tiny got an average FPS of 1.98 and 3.49 respectively. The real-time implementation was done with the help of a website to display detection images from IP cameras in real-time. In this real-time implementation, the MobileNetV2 and YOLOv4-tiny SSDs received an average FPS of 1.92 and 3.47 respectively.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	SSD MobileNet V2, YOLOv4-tiny, Convolutional Neural Network, Human Detection, Amlogic S905x
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information Technology > 59201-(S1) Undergraduate Thesis
Depositing User:	Gavin Bagus Kenzie Narain
Date Deposited:	13 Sep 2023 09:41
Last Modified:	13 Sep 2023 09:41
URI:	http://repository.its.ac.id/id/eprint/102362

Actions (login required)

View Item