Multi-Branch Inference

Because PIR and contact sensors cannot distinguish the number of people, we deal with this situation with a multi-branch inference approach to consider the context of sensors data. In some cases, the system may not be able to estimate the number of people accurately, like in the case where there are more than two candidate rooms that satisfy the room transfer condition, but after some inferences, the estimated result can finally converge to the ground-truth number. Figure 6 shows a simple example.

The maintenance of the proposed multi-branch inferring method is as follows:

**Figure 6.** Example of transfer dilemma: One person is in Room A and another one is in Room C. If the PIR sensor in Room B is activated, it is hard to determine whether the person who activated it comes from Room A or Room C, until other events occur.

• **Create Branch:** When a dilemma case occurs, several new inference branches are created for every possible movement case. Every branch has a confidence attribute representing its reliability. For example, for the transfer dilemma case in Figure 6, two inference engines are created—one for a possible transfer from Room A to Room

B; another for the transfer from Room C to Room B, with independent room status values. Both continue to infer the house status simultaneously.


$$Confidence\_{i, t \to \Delta} = \frac{1}{\max(\text{Confidence})} \times confidence\_{i \to t} \tag{3}$$

