A probabilistic exclusion principle for multiple objects

This chapter proposes a mathematically rigorous methodology for tracking multiple objects when the number of objects is fixed in advance. The fundamental problem to be addressed is demonstrated in Figure 6.1. Two instantiations of the same tracking algorithm, with different initial conditions, are used to track two targets simultaneously. When one target passes close to the other, both tracking algorithms are attracted to the single target which best fits the head-and-shoulders model being used. One might think of avoiding this problem in a number of ways: interpreting the targets as “blobs” which merge and split again (Haritaoglu et al., 1998; Intille et al., 1997), enforcing a minimum separation between targets (Rasmussen and Hager, 1998), or incorporating enough 3D geometrical information to distinguish the targets (Roller et al., 1994). However, each of these solutions can be unattractive. A blob interpretation does not maintain the identity of the targets, and is difficult to implement for moving backgrounds and for targets which are not easily segmented. A minimum separation relies on heuristics and fails if the targets overlap. Incorporating 3D information is impossible without detailed scene modelling.