الفهرس 
Introduction and SurveyOnly 14 pages are availabe for public view
 |  | from | 170 |  |  |
| | | from | 170 | | |
Abstract
Autonomous vehicles are one of the international future goals. Streets that have autonomous vehicles lead with efficiently developed software, can reduce the num- ber of accidents caused by human factors. An efficient vision system that can identify the vehicle surrounding environments (Such as vehicles, traffic signs, road signs, etc.) is essential for the success of autonomous vehicles. Vehicle detection is one of the basic elements in the autonomous vehicle vision system. A lot of at- tention has been paid for this topic, but optimization is needed to reach a reliable detection system.
The objective of this thesis is to develop and implement a reliable vehicle detection system using computer vision techniques on Graphical Processing Units (GPUs), making a comparison between a classic vehicle detection system ”Compass HOG” and a modern vehicle detection system ”YOLO V2”, making a comparison between vehicle detection systems on CPU and on GPU , implementing a GPU-based com- puter vision approach for vehicle detection by studying different techniques based on deep learning to make image classifications, in order to choose the best one that fits inside the different vehicle detection techniques that based on deep learn- ing. These CNN architectures for vehicle classifications are ResNet, InceptionRes- netV2, InceptionV3, NASNet, MobileNetV2, and PNASNet architectures. Also, there are two different datasets have been trained in these architectures to evaluate them. These datasets are Kitti dataset to train on car detection only, in additions to MIO-TCD dataset to detect various types of vehicles. For vehicle detection based on deep learning, Faster R-CNN, YOLO ”You Only Look Once”, and SSD ”Single Shot Detector” have been compared in this thesis in speed an accuracy perspectives, in order to choose the best suitable one among the others.