Recent advances in artificial intelligence (AI) have transformed fossil analysis, yet automated segmentation of dinosaur footprints remains understudied due to data scarcity and methodological limitations. This study explores a method to collect enough dinosaur footprint images for segmentation tasks, and the potential of deep learning models, particularly Mamba-based architectures, to address the unique challenges of dinosaur track segmentation. Firstly, we propose a large language model-assisted method to collect a dinosaur footprint dataset, called DinoF, containing 445 images from 72 publications. Also, we evaluate the limitations of traditional CNNs (e.g., UNet) and Transformers in handling global context and computational costs, and propose a VSLCG-U model combining spatial-state models (SSMs) with attention mechanisms. Specifically, we improve the contiguity of the scan method of SSMs, thus enhance the spatial continuity of the global information. Experiments on DinoF demonstrate best accuracy (87.63%) and reduced training time compared to existing approaches. Our work fills a critical gap in the field of automatic segmentation of dinosaur footprints.