Analysis of Lumbar Spine Radiographs

Xrays of the lumbar spine can be used by physiotherapists, orthopaedic surgeons etc to diagnose abnormalities of the lumbar spine. However, current analytic techniques such as locating the Instantaneous Centre of Rotation, require an expert to calculate and are very time consuming (1-2hrs per patient). As such, they are rarely used in practice. This project aims to speed up and ultimately automate this procedure using computer vision techniques.
 

Eye and Hand Tracking

Eye tracking has been used for some time in interactive displays, however traditional approaches require expensive and cumbersome equipment. The aim of this project is to create a system which allows real-time eye tracking with a standard PC and video camera. We are currently implementing Viola and Jones technique from CVPR2001 for locating faces in video images. We are also looking at techniques for hand-pose estimation from video sequences for use in a virtual sculpting tool. This work is part of the Watching Window project.  

Previous Research
 

Research into Teaching Computer Vision
We have conducted research into teaching computer vision using problem-based learning and primary source literature. The page linked here contains some information about that research.
 

Optical Flow Algorithm Evaluation
It has been nearly twenty years since Horn & Schunk published their influential paper on the calculation of optical flow. Since then a great deal of research has been devoted to finding more efficient and more reliable techniques for calculating optical flow. But how far have we come? In an attempt to answer this question we have developed a suite of applications which allow the generation of image sequences with ground-truth optical flow fields. Click here if you want to have a look at our results,  download an image sequence and its ground-truth motion field, or download tools so you can generate your own sequences.
 

Virtual Snakes
Active Contours or `snakes' are a relaxation-based technique used to track contours through an image sequence, or find contour-like features in an image. Previous active contour implementations have great trouble dealing with the presence of occluding objects. We describe a multiple hypotheses technique, which, coupled with texture correlation allows us to overcome some of these difficulties. Click here for an overview of the technique.
 

Feature Tracking
Feature extraction and tracking is one of the most ubiquitous operations in computer vision, yet is often overlooked. We describe a feature tracking system which fuses the predictions of a conventional correspondence tracker with a novel relaxation-based tracker. The system performs much better than either system alone - with less than one third the error rate of a conventional correspondence tracker. Click here for more.

Finding Golf Clubs
Many super athletes have benfited from computerised motion analysis, however the techniques used require expensive hardware and a lot of processing power. They also tend to be quite intrusive, disrupting the athlete's natural movement. We are curently working on an algorithm which can extract the motion of a golf-swing from an image sequence taken with a domestic quality video camera. This research is very much in the preliminary stages, however you can have a look at our progress so far.
 

Feature-Based Motion Analysis
Measurement of the 3D motion of scene features in order to extract individual objects by applying a rigid object hypothesis.