To realize the accurate estimation of the asphalt pavement density， the dielectric property test and density prediction model comparison of the asphalt mixture were carried out. Four density prediction models （CRIM model， Rayleigh model， Böttcher model and ALL model） of the asphalt mixture were derived from the electromagnetic mixing theory of the multiphase composite materials. The relative permittivity values of the asphalt mixture specimens prepared by superpave gyratory compactor and its components were measured by Percometer dielectric constant detector. Considering the gradation type and void ratio of the asphalt mixture， the errors between the density values estimated by different models and measured by surface dry method were thoroughly analysed. Finally， the density prediction accuracies of the optimal models were verified by the Ground Penetrating Radar detection and core data of the on-site AC-20 asphalt pavement. The results demonstrate that the consistency of asphalt pavement relative permittivity measurement between Percometer and Ground Penetrating Radar can be obtained. The accuracy of asphalt mixture density prediction model is affected by gradation type and void ratio. Rayleigh model and ALL model are more suitable for the density prediction of asphalt mixture than CRIM model and Böttcher model. ALL model has the highest precision.