When service robots present in environments coexist with people, human-aware navigation become an important problem to be addressed. For doing so, this study designed a human-aware motion planner by inference of human motion modes in a camera network. Given a grid map of an indoor environment, human motion mode can be defined as a probabilistic form and fused into a probabilistic grid map in order to adjust robot navigation strategies. First, a two-level learning algorithm is employed to learn motion modes of persons based on collections of trajectories which are recorded by a camera network. Subsequently, a chain of Gaussian distributions are applied to describe each motion mode. Based on these modes, human motion prediction can be computed. Experimental results show the effectiveness of fusing human motion tendency to adapt robot navigation behaviors.