This thesis concerns the evaluation of an image subtraction statistic that is used by a prototype of a chiropractic image processing package to track spinal movement. The image subtraction statistic is calculated by summing the absolute differences in pixel intensity of two images. Also included in the thesis is a brief discussion of different methods of tracking and a literature search of alternative statistics that may be appropriate for the image type (low contrast and noisy).
In summary the experimental work concluded that inter frame rotation does not have a significant effect on the performance of the image subtraction statistic when tracking inter-frame but when tracking from a particular frame to one which is significantly later in the sequence rotation must be included in the algorithm. It was also found that discretisation of the image had a detrimental effect on performance. This can be compensated for by adding a sub-pixel location calculation into the algorithm. In the original prototype a median filter (rank 5) was used to smooth the noise in the image to be searched. This was found to have marginal affect on the performance of the statistic.
Many of the algorithms presently defined in the literature were found to be unsuitable for this application as they tracked clearly defined lines or searched for a two-dimensional shape that matched a predefined three-dimensional model.
An algorithm that may prove to be a suitable alternative compared the rate of change in intensity across a window so is based on locating a change of intensity pattern rather than a pixel to pixel comparison.
There are some features that could be included in the tracking procedure to make the algorithm more efficient (the two-dimensional logarithmic search) and provide checks to safeguard against points incrementally deviating from the correct location as tracking progresses (referencing a moused frame, using the vertebra rigid body property). The benefit of incorporating the safeguard features would have to be weighed against the cost of extra computational time.
In conclusion, the image subtraction technique can be improved from, in some cases, total tracking loss to accuracy within two pixels of the correct location. This is achieved by tracking inter frame, that is from one frame to the next in the video sequence, and including a sub-pixel location calculation.