Dissertation/Thesis Abstract

Deep Learning Robust Perception Paradigm for Autonomous Vehicles
by Ali, Alaa Mohamed Leithy, Ph.D., University of Louisiana at Lafayette, 2017, 134; 10681320
Abstract (Summary)

Robust and reliable environment perception is an essential challenge to realize artificial intelligence robotics, smart traffic management, driver assistance, and autonomous driving. The present research focuses on high-performance vision based perception modules in terms of speed and precision, such as traffic sign recognition and multi-target object detection in the context of vehicle and pedestrian detection for autonomous driving. Several detection techniques are developed based on deformable part models (DPMs) and Convolutional Neural Network (CNN) for its foreseen precision. Specifically, we propose multiple approaches to improve the performance of DPM and a more accurate DeepPyramid DPM model which maps DPM into multiple CNN layers as an assembly of the two concepts with intensive computations. Multiple techniques are employed to build our frameworks, including but not limited to Fast Fourier transform, feature pyramid approximation, early classification, edge boxes region proposals, neighborhood aware cascade, and region proposal networks. In addition, our implementation employs low-level optimization techniques to achieve a faster and real-time detector, including Single Instruction Multiple Data (SIMD) instructions, multiple cores, assembly programming and, cache management. Our experimental results demonstrate that our fast DPM can process Pascal VOC 2010 at 47HZ and VGA image at 40HZ on CPU with the loss of 2% in precision, while fast DeepPyramid DPM processes PASCAL VOC at 17 fps and VGA images at 10 fps on a GPU (about 11x faster than DeepPyramid DPM). Furthermore, we introduce a programmable, low-power hardware implementation of DPM-based object detection for real-time applications. The proposed hardware circuit uses 65nm CMOS technology and consumes only 36.5mW (0.81 nJ/pixel) based on the post-layout simulation and uses the area of 3362 kgates. Finally, we present a robust end-to-end approach based on the aggregate channel features and CNN to simultaneously and effectively detect and classify traffic signs.

Indexing (document details)
Advisor: Bayoumi, Magdy
Commitee: Chu, Chee-Hung, Kumar, Ashok, Maida, Anthony
School: University of Louisiana at Lafayette
Department: Computer Engineering
School Location: United States -- Louisiana
Source: DAI-B 79/09(E), Dissertation Abstracts International
Subjects: Computer Engineering, Computer science
Keywords: Autonomous driving, Computer vision, Deep learning, Environmental perception, Object detection
Publication Number: 10681320
ISBN: 9780355854190
Copyright © 2019 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy