Development of an advanced system on GPU platforms for understanding the environment of the autonomous vehicles

Abstract 

The problem we are dealing with in the diploma thesis is the perception of important objects in pictures and video of driving on Slovenian roads. To solve this problem, we decided to use machine learning techniques and a framework Darknet for training neural networks. We created our own set of data or. images and use Darknet to train the neural network model. We tested this model on powerful hardware, and on the NVIDIA Jetson TX2 edge platform.

NVIDIA Jetson TX2 is edge GPU microcomputer. It is based on the Tegra X2 architecture, which contains 2 central processing units: Denver2 and Cortex-A57. These together contain 6 process cores. Jetson supports up to 8GB of read-write memory. It features Pascal GPE architecture with a processing power of up to 750 GFLOPS at 32-bit floating-point numbers. Power consumption is up to 15W [9]. The Linux distribution is usually used for the operating system, nVidia offers a ready-made image of the Ubuntu 16.04 distribution. The operating system offers us CUDA driver support for the GPU. CUDA support is critical as it allows us to use tools and software that offer machine learning capabilities.  

Darknet is an open source framework for deep neural networks. It is written in the C, C ++ and CUDA programming languages. By the name Darknet, many people think of a part of the web that is not accessible with ordinary software. However, in this case, we are not talking about this topics, but about the software of creator Joseph Redmon, whose works are available on his GitHub page pjreddie. Software names are ambiguous and possibly provocative. As it turns out, this is done purposefully and is a common theme of this creator when naming programs and components. The software is nevertheless professional, and the algorithms used in this software achieve good results in the field of machine learning.

YOLO is a neural network capable of detecting and classifying objects in images [13]. The name YOLO comes from the phrase “you only look once”, which means “you only look once”. During the development, the authors created a third version, called YOLOv3, and a fourth version was released at the time of writing. Each version is an upgrade of the previous one. When the authors released the first version, they raised a lot of dust among neural network researchers. YOLO was significantly faster than other neural networks for object detection and classification. It was so fast that it was able to detect objects in video in real time and with acceptable accuracy using a more powerful GPU.