Abstract: To implement the object-based video coding, a video sequence is needed to segment it into several individual objects. A spatiotemporal algorithm is therefore presented for automatic video segmentation. In the temporal state a preliminary segmentation is done through detecting the difference between two successive frames to locate the moving object in rough, while in the spatial state the predictable watershed algorithm is used to accelerate the conventional one and make the temporal segmentation result more exact. The two results in both temporal and spatial states are complementary to each other to benefit the automatic video segmentation. Furthermore, the whole segmentation process is performed on the images after wavelet transform to avoid oversegmenting. Experimental results showed that the proposed method provides not only the better segmentation quality but also the low computational cost, especially the more exact and practical semantic information.