Although there are many techniques for determining image threshold values for binarising intensity images [16] most of them make assumptions that do not hold for standard difference images, thereby invalidating their application. For instance, many methods expect the intensity histogram to be bimodal, and some cannot cope if the modes are too dissimilar in size, or if they are not approximately Normal. We know of rather few thresholding techniques specifically designed to be effective for difference maps; these are described below.
Yang and Levine [19] employed a two stage process to select local and global thresholds for differenced edge maps using the full image sequence. First, for each edgel, the median absolute deviation ( ) of all the edgels at that location over the sequence is calculated. The edgel's local difference map threshold is then set to . Next, a histogram is formed from those edgels which are above their local threshold values. The median absolute deviation calculated over the histogram is used to generate a global threshold . Edgels now are marked as change only if they exceed both their local and the global thresholds.
Jain and Nagel [8], and later Hsu et al. [6], bypassed the difference map and detected change using hypothesis testing. The images were modelled as patches whose intensities were described by bivariate polynomials. For each pixel a likelihood test checked whether the intensities within a local window in each of the two images could have been drawn from a single intensity distribution. Since the t-test is used the threshold can then be related to a confidence level.