The complex environment of pig farms and the dynamic growth of pigs, leading to changes in body shape, pose challenges for accurate pig segmentation. Video object segmentation, with its ability to utilize temporal information from video sequences, offers a promising solution for handling dynamic objects. In this study, we focus on pigs during dynamic feeding and growth processes in performance testing, constructing a pig video dataset comprising 234 video sequences. We propose a semi-supervised pig video segmentation method based on XMem-SimAM. By introducing SimAM attention for multi-scale feature fusion, the model"s ability to extract temporal information at different scales is enhanced, capturing the temporal characteristics of pigs" dynamic movements. The spatial-channel attention module is employed to strengthen the model"s extraction of temporal semantic feature weights. By optimizing the multi-scale feature fusion strategy and the upsampling module, the temporal correlation information in video sequences is fully utilized, improving the segmentation accuracy of pigs in videos at a fine-grained level. Comparative tests show that the XMem-SimAM model achieves a Jaccard index of 96.9, contour accuracy F-score of 95.8, average metric J F of 98.0, and Dice coefficient of 98.0 on the pig video dataset, outperforming video object segmentation methods such as MiVOS, STCN, DEVA, and XMem++. In the inference phase, the processing speed reaches 58.5FPS, with a memory consumption of 795MB, achieving a good balance between processing efficiency and resource utilization. The results indicate that this study provides a new method for segmenting dynamically growing pigs in complex farm environments.