where  c   is a constant  representing  the influence  of

the current image in each ground image update,  and c3 is a constant representing the frequency at which new ground images are produced。 Both these parameters are obtained in a heuristic fashion or through the decisions of higher-level  processes。

2。4。 Figure  image segmentation

Once a figure image has been obtained, we can consider the task of using the information that it provides to identify objects of interest。 Unfortu- nately, figure images tend to contain pixels that belong to a variety of items other  than  just  the objects of interest。 For example,  one may  find that the figure image indicates activity caused by shadows, camera jitter, or salt-and-pepper noise。 Another problem that appears with the analysis of figure images is that an image may contain multiple objects of interest。 In order to distinguish better among several objects, the figure image is partitioned into segments each of which represents a single object。 This figure image segmentation has a beneficial side- effect of identifying and removing many instances of the problem cases described  above。

The figure image segmentation is achieved through a  single  pass  of  a  sequential  labeling  algorithm 25

Because the algorithm creates segments from only a single pass through the figure image, all statistics that are to be calculated for the segments  must be obtained dynamically。 The selection of which statistics to calculate depends on the segment analysis (see Section 3。 I)。 In some cases, it may suffice to maintain a size (pixel count) and minimum bounding rectangle。

By  following  the  previous   algorith   25  without

further modification, there can be as many as several hundred segments generated in a single image, only a few of which describe actual objects of interest。 Since the computational performance of object detection relies on keeping the number of considered segments to a minimum, modifications must be made in  order to make the process more efbcient。 After each scanline is processed, an analysis is made of all segments that have been completely processed by previous scanlines。 If any of these  completed segments are found to have  different  dimensions than those of a typical object of interest,  then  they are removed from the segment collection data structure。 In practice, this pruning removes a large number of the undesirable sources of figure segments。 By continuously minimizing the number of segments in the data structure, this process also improves the speed at which  the segmentation  can occur。

(a) (b)

                                                                                   (d)

(a) A manipulator has positioned a camera looking down upon a wooden platform。 The system has initially stored a ground image in memory。

(b) After a short period of time, a target (a remotely controlled toy car) appears under- neath the robot。

(c) A current image of the scene with a moving target。

(d) The target becomes readily apparent, as a figure image is formed by comparing the current and the ground images。

Fig。  1。  Construction  of  a  figure image。