As challenges such as serious occlusions and deformations coexist, video segmentation with accurate robustness has become one of the hot topics in computer vision. This study proposes a video segmentation method with absorbing Markov chains and skeleton mapping, which progressively produces accurate object contours through the process of pre-segmentation—optimization—improvement. In the phase of pre-segmentation, based on the twin network and the region proposal network, the study obtains regions of interest for objects, constructs the absorbing Markov chains of superpixels in these regions, and calculates the labels of foreground/background of the superpixels. The absorbing Markov chains can perceive and propagate the object features flexibly and effectively and preliminarily pre-segment the target object from the complex scene. In the phase of optimization, the study designs the short-term and long-term spatial-temporal cue models to obtain the short-term variation and the long-term feature of the object, so as to optimize superpixel labels and reduce errors caused by similar objects and noise. In the phase of improvement, to reduce the artifacts and discontinuities of optimization results, this study proposes an automatic generation algorithm for foreground/background skeleton based on superpixel labels and positions and constructs a skeleton mapping network based on encoding and decoding, so as to learn the pixel-level object contour and finally obtain accurate video segmentation results. Many experiments on standard datasets show that the proposed method is superior to the existing mainstream video segmentation methods and can produce segmentation results with higher region similarity and contour accuracy.