| Skip to main content | Skip to navigation | Skip to Search |

[University home]

Research Computing

Michael May

me.jpg

Information

Room 1.014
Research Computing Services
Devonshire House
The University of Manchester
M13 9PL

0161 306 6644
michael.may@postgrad.manchester.ac.uk

Mendeley Profile

PhD Blog

Introduction

I am a PhD student within the School of Computer Science and work out of Research Computing Services. My project is entitled Image Processing for Unmanned Air Vehicle Applications. I began in October 2008. Key people associated with the project are:

Publications

2012

2010

2009

Academic Qualifications

Image Processing for Unmanned Air Vehicle Applications

Background

Unmanned air vehicles (UAVs) are being increasingly used for both civil and military operations where vehicle size, risk or cost precludes the use of manned systems. A key enabling technology for UAVs operating close to the ground or in urban environments is the use of computer vision to determine the position of the UAV with respect to objects and to recognise particular objects that may be of interest.

The overall aim of the PhD project is to develop a robust, real time visual object recognition system using video data from a small autonomous air vehicle. The proposed algorithm for image processing is SIFT (Scale Invariant Feature Transform). It is proposed to use the parallel graphics processing hardware in commercial PC graphics cards to enable real time capability. The work will be innovative in that the existing SIFT algorithm will be adapted to the specific problem of real time object recognition in cluttered urban environments.

SIFT Technology and GPGPUs

In the few years since it was developed, the SIFT image recognition model has been used extensively, being considered the most robust method available in the image processing community. One of the main drawbacks of SIFT algorithms is that they are very computationally expensive.

Thanks to enormous funding into the computer games industry, massively parallel (90+ core) graphics processing units are now available at low cost and very recently in portable devices. When combined with open source graphics code libraries, we have the ability to run very demanding algorithms in real time via GPGPU (General Purpose computation on GPUs) techniques. Shadowfax is the primary machine used for this purpose. The project aims to develop parallel algorithms which address these problems, emphasising their value in the growing research area of UAV guidance and situational awareness.

Progress will initially stem from gathering a greater understanding of SIFT and its failings as a recognition algorithm. Initial interest/projects areas cover the parallelisation of SIFT and the response of the SIFT algorithm when used in conjunction with high dynamic range images.

SIFT Parameter Sweep

The parameters of SIFT affect the response of the algorithm but exactly how changes in them vary the result and accuracy of feature matching has not previously been studied in detail. There is currently no intelligent means of setting the pre-processing parameters which the algorithm uses to decide whether an image feature is worth tracking or not. This is one of the main areas which I am studying and a paper has been written to show the results of the sweeps across the SIFT parameters using a GPU implementation of SIFT. It has been accepted by the BMVC 2010 PhD workshop for presentation (see publications).

The following files contain the code, data and results from the experiments (last updated 4/8/2010):

GPU SIFT Parameter Sweep code and images.

GPU SIFT Parameter Sweep data.

GPU SIFT Parameter Sweep Matlab histogram script.

Xmdv tool has been used for data analysis.

Video

Marten Bjorkman's Original Cuda Sift

SIFT Feature Weighting

Further work analysing and improving SIFT has led to the development of a novel, 3D-based, SIFT weighting technique which utilises the 3D data from a pair of stereo images to cluster and class matched SIFT features. Weightings are applied to the matches based on the 3D properties of the features and how they cluster to attempt to discriminate between correct and incorrect matches. Progress was made in further understanding the noise level (percentage of false positives) that commonly occurs with the SIFT algorithm which allows the features to be weighted. This has been achieved by analysing large sets of randomly generated features which are known not to correctly match to a scene. By calculating the percentage of matches to those which are known to be relevant to the scene. It has been shown that approximately 3 percent of the random features match. The match ratio of total features in the target image to the number of matched features can be used to calculate whether the matches are be due to noise or if they are statistically significant. Therefore 3 percent is a minimum threshold for successful object detection.

By divding the features into three different types using the 3D information it has been shown that the signal to noise ratio varies across different feature types and that that the type 3 3D features have a higher likelihood of being correct when compared to a randomly generated feature set. Using this noise information the features can be weighted accordingly to determine which matched features are more reliable. This allows a threshold to be used to select the best matches from a scene and discard the mismatches.

High Dynamic Range SIFT Feature Extraction

This work introduces a process where fusion features assist matching SIFT image features from high contrast scenes. FAW defines the order for extracting features: features, alignment then weighting. The process uses three quality measures to select features from a series of images, each exposed differently, and weight the features in favour of those areas that are defined as well exposed. The results show an advantage in using these features over features extracted from the common alternative techniques of exposure fusion and tone mapping which extract the features as AWF; alignment, weighting and features. This work has also shown that the process allows for a more robust response when using misaligned or stereoscopic image sets.

Thesis Chapters - Work In Progress

Current Draft Automatically Updated (Dropbox Link)

Page updated on: 24/03/2012