Degree Type

Thesis

Date of Award

2015

Degree Name

Master of Science

Department

Electrical and Computer Engineering

First Advisor

Arun K. Somani

Abstract

Detection of objects of interest is a fundamental problem in computer vision. Foreign object detection (FOD) is to detect the objects that are not expected to be appear in certain area. For this task, we

need to first detect the position of foreign objects, and then compute the distance to the foreign objects

to judge whether the objects are within the dangerous zone or not. The three principle sources of difficulty in performing this task are: a) the huge number of foreign objects categories, b) the calculation of

distance using camera(s), and c) the real-time system running performance. Most state-of-art detectors

focus on one type or one class of objects. To the best of our knowledge, there is no single solution that

focuses on a set of multiple foreign objects detection in an integrated manner. In some cases, multiple

detectors can operate simultaneously to detect objects of interest in a given input. This is not efficient.

The goal of our research is to focus on detection of a set of objects identified as foreign object in an

integrated and efficient manner. We design a multi-class detector. Our approach is to use a coarse-tofine strategy in which we divide the complicated space into finer and finer sub-spaces. For this purpose,

data-driven clustering algorithm is implemented to gather similar foreign objects samples, and then an

extended vector boosting algorithm is developed to train our multi-class classifier. The purpose of the

extended vector boosting algorithm is to separate all foreign objects from background. For the task of

estimation of the distance to the foreign objects, we design a look-up table which is based on the area

of the detected foreign objects.

Furthermore, we design a FOD framework. Our approach is to use stereo matching algorithm to

get the disparity information based on intensity images from stereo cameras, and then using the camera

model to retrieve the distance information. The distance calculated using disparity is more accurate

than using the distance look-up table. We calculate the initial distance map when no objects are in the

scene. Block of interest (BOI) is the area where distance is smaller than the corresponding area in the

initial distance map. For the purpose of detecting foreign objects, we use flood fill method along with

noise suppression method to combine adjacent BOI with higher confidence level.The foreign object detection prototype system has been implemented and evaluated on a number of

test sets under real working scenarios. The experimental results show that our algorithm and framework

are efficient and robust.

DOI

https://doi.org/10.31274/etd-180810-4113

Copyright Owner

Haoyuan Lin

Language

en

File Format

application/pdf

File Size

39 pages

Share

COinS