Date on Master's Thesis/Doctoral Dissertation

5-2017

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Computer Engineering and Computer Science

Degree Program

Computer Science and Engineering, PhD

Committee Chair

Frigui, Hichem

Committee Member

Amini, Amir

Committee Member

Nasraoui, Olfa

Committee Member

Park, Juw Won

Committee Member

Zhang, Hui

Author's Keywords

vehicle recognition; multiple instance learning; saliency detection; deep learning; fine-grained classification; automated surveillance

Abstract

Vehicle Make and Model Recognition (VMMR) has evolved into a significant subject of study due to its importance in numerous Intelligent Transportation Systems (ITS), such as autonomous navigation, traffic analysis, traffic surveillance and security systems. A highly accurate and real-time VMMR system significantly reduces the overhead cost of resources otherwise required. The VMMR problem is a multi-class classification task with a peculiar set of issues and challenges like multiplicity, inter- and intra-make ambiguity among various vehicles makes and models, which need to be solved in an efficient and reliable manner to achieve a highly robust VMMR system. In this dissertation, facing the growing importance of make and model recognition of vehicles, we present a VMMR system that provides very high accuracy rates and is robust to several challenges. We demonstrate that the VMMR problem can be addressed by locating discriminative parts where the most significant appearance variations occur in each category, and learning expressive appearance descriptors. Given these insights, we consider two data driven frameworks: a Multiple-Instance Learning-based (MIL) system using hand-crafted features and an extended application of deep neural networks using MIL. Our approach requires only image level class labels, and the discriminative parts of each target class are selected in a fully unsupervised manner without any use of part annotations or segmentation masks, which may be costly to obtain. This advantage makes our system more intelligent, scalable, and applicable to other fine-grained recognition tasks. We constructed a dataset with 291,752 images representing 9,170 different vehicles to validate and evaluate our approach. Experimental results demonstrate that the localization of parts and distinguishing their discriminative powers for categorization improve the performance of fine-grained categorization. Extensive experiments conducted using our approaches yield superior results for images that were occluded, under low illumination, partial camera views, or even non-frontal views, available in our real-world VMMR dataset. The approaches presented herewith provide a highly accurate VMMR system for rea-ltime applications in realistic environments.\\ We also validate our system with a significant application of VMMR to ITS that involves automated vehicular surveillance. We show that our application can provide law inforcement agencies with efficient tools to search for a specific vehicle type, make, or model, and to track the path of a given vehicle using the position of multiple cameras.

Share

COinS