When extracting vehicle vibration acceleration from smart phone through software developed, the data quality of the vibration acceleration of the vehicle should be guaranteed in order to correctly evaluate the track smoothness and vehicle running comfort. This paper establishes an abnormal value recognition model based on the methods of probability-statistics and wavelet, in which median filter and wavelet filter are used to eliminate the random errors caused by the performance stability of mobile sensors and the change of testing environment. The effect of two filtering methods on the random error of mobile phone detection data is verified by combining with the detection data from Chengdu metro. The analysis result shows that the abnormal value location of mobile phone detection data can be accurately extracted based on the abnormal value recognition model. The mobile phone detection data truly reflect the vibration response of the car body by eliminating random errors caused by external environmental changes using methods of median filter and wavelet filter, so the model can be used to correctly evaluate the track smooth state and vehicle operating comfort.