An improved method that uses 3D-DWT-SPIHT in original video group and residual error group alternatively was proposed. In addition, the multi-wavelet bases were combined in time domain. The proposed method reduces coding latency time and releases memory requirement, and improves both compression efficiency and reconstructive image's quality, especially to the video sequences with complicated background. Simulation results show that, in the case of equal compression ratio and memory space, our method has higher quality(PSNR above 1dB) than available methods. It is still slightly better than others', when memory is reduced to half.