Year

2015

Degree Name

Doctor of Philosophy

Department

School of Civil, Mining and Environmental Engineering - Faculty of Engineering and Information Sciences

Abstract

Visually impaired people facemany difficulties in daily life. Among their essential activities, traveling safely and independently is one of themost challenging tasks. This thesis focuses on assisting vision-disabled people to deal with this problem. We propose a non-invasive system that locates various types of walking lanes for visually impaired travelers, using a novel image processing architecture and machine learning algorithms. In this system, a camera is employed to capture the scene image in front of the traveler. Then, the lane type is identified as markedlanes, unmarked-lanes or non-lanes. Finally, a suitable lane detector for the lane type is applied to locate the walking region.

To recognize the lane type in each image, we propose a new method using multiple instance learning. The proposed method represents each image as a bag of instances. Each instance is an image region and described by a feature vector. A vocabulary-based framework of multiple instance learning is then employed to categorize bags.

To detectmarked-lanes that are located by lanemarkers, we propose a regionbased method. The proposed method for marked-lane detection extracts first local image regions locating on the borders of lanemarkers. The lanemarkers are then found using a Markov random field model. The walking region is finally determined using the geometric cues of lane marker pairs.

We also propose a new method to find pedestrian lanes in unstructured environments where lanes have no painted markers, vary significantly in appearance and have different shapes. The proposed method for unmarked-lane detection locates the walking lane using both appearance and shape features. The lane ap- pearancemodel is learned on-the-fly froma sample region,which is automatically determined employing the vanishing point and the properties of lane surfaces and boundaries. Shape contexts are employed tomodel the shape of pedestrian lanes.

All the proposed methods for classifying and detecting pedestrian lanes are evaluated on a large and new data set of images, collected from different scenes under challenging illumination conditions. The experimental results prove the robustness and efficiency of the proposed methods.

FoR codes (2008)

0801 ARTIFICIAL INTELLIGENCE AND IMAGE PROCESSING

Share

COinS
 

Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.