One of the critical research issues for the future robot is the avoidance of collisions with people while it moves among them. Therefore, people-tracking and motion prediction are important. People are tracked more readily and successfully by using distributed sensor networks than robot's local sensors. Floor sensor networks, in particular, are resistant to changes in lighting conditions and other environmental disturbances. However, the problem with the system is that a person walking is observed as if it were being done in "rabbit hops": The signal is nonlinear and even nonholonomic. We tackled the problem by assuming that human walking is regular in terms of walking rhythms. Then the signal was modeled by an oscillation model based on four walking parameters: cycles, phases, directions, and strides. To adapt the parameters to irregular walking, the model was multi-hypothesized based upon a particle filtering algorithm. Experimental results showed that the multi-hypothesized oscillation models showed more than 80% tracking accuracy for four walking patterns: straight walking, stop-and-go, turn, and winding walking. Further, the models were superior to the nearest neighbor filter with regard to the performance of data association for two persons.