A multiple object tracking algorithm, for tracking passengers boarding and exiting a bus, is described in this paper. The system uses a single visible light camera, and is required to run in real time with commercially available hardware. Thus, computational efficiency is one of its major requirements. The authors claim that many existing algorithms are not applicable here, due to the proximity of the targets to the camera, and the substantial variations of operating environments.
The authors tackle the problem by, first, taking the absolute difference of two consecutive images; second, applying a simple threshold to the absolute difference images; third, partitioning the threshold image into hexagonal cells with six neighbors; and, fourth, carefully applying a set of rules to condition the threshold cells. The conditioning will take care of a number of malformations, including lakes (holes), peninsulas, straits, bays, isthmuses, and merged multiple targets, which result from the simple thresholding and temporal differencing.
Mabey and Gunther have carefully crafted procedures to handle each malformation, which appear to work well for this particular application. However, by doing so, they introduced many parameters that needed to be adjusted. For this reason, I think that it is not easy to apply the algorithm to different tasks and environments. More experimental results and comparisons would be appreciated.
]]