AFM based nanomanipulation has becoming a promising approach in developing devices and structures at nanoscale. One of the prerequisites for the effective and successful nanomanipulation is that the relative position between the AFM tip and the manipulated object can be controlled accurately. However this prerequisite is still hindered by the spatial uncertainties aroused from the PZT nonlinearity and thermal drift. This paper proposes a probability referenced tip localization method to solve these problems. The key idea is that the tip position is not only calculated from the applied PZT voltages, but also combined with the sensing information about the landmarks set in the sample surface. Furthermore, the uncertainties in the process of both tip motion and landmark sensing are described and considered in tip localization, which results in a higher positioning accuracy. In addition, the landmarks are selected from the features of sample surface, positioning error due to the thermal drift can be compensated automatically. Both the simulation and experimental results are presented to demonstrate the effectiveness and efficiency of the proposed method.