To slove the problem of high taxi empty-loading ratio of taxi and difficulty in finding passengers, a new model called Possion-Kalman combined prediction Model (PKCPM) was proposed. Firstly, weighted Non-Homogeneous Poisson Model (NHPM) was used to get the estimated value of the target time based on taxi historical data. Secondly, the mean value of the passenger demand in the near time, was taken as the predicted value, based on the real-time data. Finally, the predicted value and the estimated value were used as the inputs of Kalman filtering model to predict the target variance, meanwhile, the error backpropagation mechanism was introduced to reduce the next prediction error. The experimental results on the taxi trajectory dataset in Zhengzhou show that compared with NHPM, Weighted NHPM (WNHPM) and Support Vector Machine (SVM), PKCPM achieves a better optimization effect, and the error of PKCPM is reduced by about 8.85 percentage points and 14.9 percentage points respectively compared with WNHPM and SVM. PKCPM can predict passenger demand within different time and spacial grid, and provides a reliable solution to taxi driver for finding passengers.