1 Introduction

Due to its simplicity image differencing is a popular method for change detection. It only requires calculating the absolute values of the difference between the corresponding pixels in two images, and large values in the difference map then indicate locations of change. Common applications of image differencing include object tracking [ 25 ], intruder surveillance systems [ 3 , 5 ], vehicle surveillance systems [ 7 , 8 , 11 ], and interframe data compression [ 2 ]. There are also many examples of its use for analysing satellite images [ 20 ] to measure land erosion, deforestation, urban growth, crop development, etc., and for analysing medical images to measure cell distribution [ 10 ], etc.

The difference map is usually binarised by thresholding it at some pre-determined value to obtain a change/no-change classification. However, the threshold value is critical, since too low a value will swamp the difference map with spurious changes, while too high a value will suppress significant changes. The proper value of the threshold is dependent on the scene, possibly fluctuating camera levels, as well as viewing conditions (e.g. illumination) which may change over time. This indicates that in general the threshold value should be calculated dynamically based on the image content, and that experimentally selecting a value (e.g. Jain [ 8 ], Koller et al.  [ 11 ]) is not appropriate for a robust autonomous vision system.

As an extension to global threshold determination there are various other procedures that can improve change detection. Local thresholding can be useful, particularly when the scene illumination varies locally over time. Noisy difference maps can be much improved by removing small isolated change pixels, merging close regions of change, incorporating connectivity, and performing hysteresis thresholding [ 1 , 11 , 18 , 25 ]. Rather than differencing adjacent frames in temporal image sequences background images can be dynamically generated [ 12 , 18 , 25 ], and these are differenced with each image instead. However, these issues will not be further explored in this paper.

Several general approaches are possible for determining thresholds for change detection. First the signal, noise, or both can be modelled. Second, either their intensity and/or spatial properties can be modelled. In this paper we consider all of the four combinations of pairs of object and property (signal/noise and intensity/spatial distribution) and describe four techniques for threshold determination that fall into each of these categories (see table  1 ). Note that most standard intensity image thresholding techniques belong to the signal/intensity class.