Person tracking in videos is crucial in different areas such as security applications. In this work we present a method that first finds human presence probabilities on discrete locations via variational Bayesian inference using images obtained from omnidirectional cameras and then uses that information to solve the tracking problem as a flow optimization problem. In our experiments on the BOMNI dataset, we have increased tracking performance (MOTA) to %86.39, which was reported as %68.18 using the baseline method.