With　the　widespread　of　surveillance　video　application,　anomaly　detection　in　surveillance　video　is　significant　for　safety　maintenance.　In　this　paper,　we　propose　a　new　abnormal　detection　framework,　which　contains　an　improved　future　frame　prediction　model　based　on　GAN　network　to　locate　the　abnormal　events　in　the　video.　Unlike　the　previous　U-Net　network　which　is　usually　used　as　prediction　model　in　abnormal　event　detection,　3D　U-Net　is　adopted　in　our　proposed　network　for　predicting　future　frames　which　can　fuse　the　spatial　and　temporal　features　simultaneously　by　using　3-dimensional　convolution.　For　normal　events,　the　improved　prediction　network　takes　into　account　the　temporal　features,　resulting　in　low　abnormal　scores,　while　abnormal　events　have　high　abnormal　scores.　Experiments　show　that　this　method　can　get　good　results　on　the　Avenue,　UCSD　Ped2　and　ShanghaiTech　datasets,　and　the　AUC　values　on　these　datasets　could　reach　86.0%,　96.3%　and　73.6%　respectively.　©　2021　ACM.