Implementasi dan Optimasi Human detection Pada NVIDIA DGX A100

Gunasatwika, I Gde Ardha Semaranatha (2023) Implementasi dan Optimasi Human detection Pada NVIDIA DGX A100. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 05311940000034-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2025. Download (13MB) | Request a copy

Abstract

Perkembangan teknologi human detection menjadi tren hangat di dunia teknologi informasi saat ini, karena aplikasinya yang luas di berbagai bidang seperti otomatisasi, keamanan, dan hiburan. Namun, pengembangannya menghadapi beberapa tantangan, terutama dalam pengolahan data yang besar dan kompleks. Untuk mengatasi masalah ini, dibutuhkan GPU dengan kapabilitas yang besar seperti NVIDIA DGX A100. Pada penelitian ini, akan dilakukan uji coba untuk mengimplementasikan dan mengoptimalkan teknologi human detection pada NVIDIA DGX A100 dengan menggunakan empat algoritma Deep Learning yang terdiri dari dua buah algoritma two-stage yaitu Mask R-CNN dan Cascade R-CNN serta dua buah algoritma one-stage yaitu RetinaNet dan YOLOv8. Untuk menguji performa dari keempat model tersebut, digunakan dua jenis Compute Type NVIDIA DGX A100, tipe 20g dengan 12 CPU Core, 60 GB CPU RAM, 3 GPU Compute, dan 20 GB GPU RAM serta tipe 40g dengan 28 CPU Core, 120 GB CPU RAM, 7 GPU Compute, dan 40 GB GPU RAM. Perbandingan dilakukan berdasarkan Average Precision (AP) dan Frame Per Second (FPS) antara keempat algoritma tersebut. Dalam proses perbandingan didapatkan bahwa performa terbaik diperoleh oleh model YOLOv8x dengan nilai AP sebesar 63,64%, FPS rata-rata pada Compute Type 20g sebesar 56,23 dan 60,97 pada Compute Type 40g. Guna mengoptimalkan performa dari model YOLOv8x, dilakukan proses training menggunakan dataset COCO-2017 yang telah di-filter menjadi hanya kelas person dengan total gambar 64,115 untuk training dan 2693 untuk validation. Hasil optimasi menunjukkan terdapat peningkatkan pada performa model, akurasi model YOLOv8x yang awalnya AP50 83,86%, AP75 69,36%, AP50−95 63,64% menjadi AP50 85,13%, AP75 69,56%, AP50−95 63,83%. Kemudian, untuk rata-rata FPS naik dari 56,23 menjadi 57,33 pada Compute Type 20g dan naik dari 60,97 menjadi 61,35 pada Compute Type 40g. Pada tahap implementasi model YOLOv8x yang telah dioptimasi pada kamera RTSP dengan spesifikasi kamera 1920x1080p 15 FPS dengan website dashboard yang telah dibuat didapatkan FPS rata-rata 13 FPS.
===============================================================================================================================
The development of human detection technology has become a hot trend in the world of information technology today, due to its wide applications in various fields such as automation, security, and entertainment. However, its development faces several challenges, especially in processing large and complex data. To overcome these problems, a GPU with significant capabilities like the NVIDIA DGX A100 is required. In this study, experiments will be conducted to implement and optimize human detection technology on the NVIDIA DGX A100 using four deep learning algorithms, consisting of two two-stage algorithms, namely Mask R-CNN and Cascade R-CNN, as well as two one-stage algorithms, namely RetinaNet and YOLOv8. To test the performance of these four models, two types of NVIDIA DGX A100 Compute Types are used. The first type is 20g, which has 12 CPU Cores, 60 GB CPU RAM, 3 GPU Compute, and 20 GB GPU RAM. The second type is 40g, which has 28 CPU Cores, 120 GB CPU RAM, 7 GPU Compute, and 40 GB GPU RAM. A comparison is made based on the Average Precision (AP) and Frames Per Second (FPS) between the four algorithms. In the comparison process, it is found that the best performance is achieved by the YOLOv8x model, with an AP value of 63.64%, an average FPS of 56.23 on Compute Type 20g, and 60.97 on Compute Type 40g. To optimize the performance of the YOLOv8x model, a training process is performed using the COCO-2017 dataset, filtered to include only the ”person” class, with a total of 64,115 images for training and 2,693 for validation. The optimization results show an improvement in the model’s performance, with the initial accuracy of YOLOv8x model being AP50 83.86%, AP75 69.36%, AP50−95 63.64%, and after optimization, it becomes AP50 85.13%, AP75 69.56%, AP50−95 63.83%. Furthermore, the average FPS increases from 56.23 to 57.33 on Compute Type 20g and from 60.97 to 61.35 on Compute Type 40g. In the implementation stage of the optimized YOLOv8x model on an RTSP camera with a camera specification of 1920x1080p and 15 FPS, using a created website dashboard, an average FPS of 13 FPS is obtained.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Human Detection, Deep Learning, NVIDIA DGX A100, YOLOv8x
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence T Technology > T Technology (General)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information Technology > 59201-(S1) Undergraduate Thesis
Depositing User:	I Gde Ardha Semaranatha Gunasatwika
Date Deposited:	25 Aug 2023 03:08
Last Modified:	25 Aug 2023 03:08
URI:	http://repository.its.ac.id/id/eprint/101765

Actions (login required)

View Item