|
" An Experimental Evaluation of Probabilistic Deep Networks for Real-time Traffic Scene Representation using Graphical Processing Units "
El-Shaer, Mennat Allah Ahmed Mohammed
Ozguner, Fusun
| Document Type
|
:
|
Latin Dissertation
|
| Language of Document
|
:
|
English
|
| Record Number
|
:
|
1105778
|
| Doc. No
|
:
|
TLpq2323168818
|
| Main Entry
|
:
|
El-Shaer, Mennat Allah Ahmed Mohammed
|
|
|
:
|
Ozguner, Fusun
|
| Title & Author
|
:
|
An Experimental Evaluation of Probabilistic Deep Networks for Real-time Traffic Scene Representation using Graphical Processing Units\ El-Shaer, Mennat Allah Ahmed MohammedOzguner, Fusun
|
| College
|
:
|
The Ohio State University
|
| Date
|
:
|
2019
|
| student score
|
:
|
2019
|
| Degree
|
:
|
Ph.D.
|
| Page No
|
:
|
107
|
| Abstract
|
:
|
The problem of scene understanding and environment perception has been an important one in robotics research, however existing solutions applied in current Advanced Driving Assistance systems (ADAS) are not robust enough to ensure the safety of traffic participants. ADAS development begins with sensor data collection and algorithms that can interpret that data to guide the intelligent vehicle's control decisions. Much work has been done to extract information from camera based image sensors, however most solutions require hand-designed features that usually break down under different lighting and weather conditions. Urban traffic scenes, in particular, present a challenge to vision perception systems due to the dynamic interactions among participants whether they are pedestrians, bicyclists, or other vehicles. Object detection deep learning models have proved successful in classifying or identifying objects on the road, but do not allow for the probabilistic reasoning and learning that traffic situations require. Deep Generative Models that learn the data distribution of training sets are capable of generating samples from the trained model that better represent sensory data, which leads to better feature representations and eventually better perception systems. Learning such models is computationally intensive so we decide to utilize Graphics Processing chips designed for vision processing. In this thesis, we present a small image dataset collected from different types of busy intersections on a university campus along with our CUDA implementations of training a Restricted Boltzmann Machine on NVIDIA GTX1080 GPU, and its generative sampling inference on an NVIDIA Tegra X1 SoC module. We demonstrate the sampling capability of a simple unsupervised network trained on a subset of the dataset, along with proling results from experiments done on the Jetson TX1 platform. We also include a quantitative study of different GPU optimization techniques performed on the Jetson TX1.
|
| Subject
|
:
|
Artificial intelligence
|
|
|
:
|
Computer engineering
|
|
|
:
|
Computer science
|
|
|
:
|
Electrical engineering
|
|
|
:
|
Robotics
|
| |