Video anomaly detection aims to identify abnormal events deviating from the expected behavior. Because the definition of abnormal and normal behavior is challenging to distinguish directly, and the occurrence of abnormal behavior is rare and random, video anomaly detection becomes challenging. This paper proposes a hybrid architecture MAU-VAD based on the premise that the semantic consistency of appearance features and motion information is high in normal events but low in anomaly events. First, we design a motion feature reconstruction module to reconstruct optical flow, which aims to record the stable motion information in normal events. Afterward, the motion information of the reconstructed optical flow and the appearance information of the corresponding original frame are extracted separately using a two-stream autoencoder, and the extracted motion features and appearance features are constrained and fused to generate predicted future frames. Since the reconstruction quality of optical flow directly affects the generation quality of final future frames, the reconstructed optical flow of abnormal events has a greater impact on the quality of generated future frames, thus improving the efficiency of anomaly detection. Experimental results on three standard public datasets demonstrate the effectiveness of the method.