In cellular network, Non-Line-of-Sight(NLOS) error is the key factor of affecting positioning accuracy. Based on kalman filter and Greenstein model, this paper identifies NLOS error by judgment and directly mitigates it with its mean modifying new abnormal information caused during kalman prediction process, then eliminates measurement error with polynomial smoothing, thus the Time-of-Arrival(TOA) measurements are reconstructed; after that, it modifies the measuring matrix and gets the locating results using the reconstructed measurements. The simulation results show that the proposed method can mitigate NLOS error effectively and improve the location accuracy in the condition of NLOS propagation, meets the demands of E-911 to some extent.