Abstract: Aimed at the low efficiency of manual comparison of tire tracks on roller tracks in the scene investigation, a few-shot semantic segmentation algorithm for small sample tire tracks based on meta-learning is proposed. A pre-trained feature extractor was used to map support images to the depth feature space, and the adaptive target classifier was constructed based on vision transformer (ViT). The whole model was optimized using the reinforcement learning in the fine-tuned manner. The results show that the proposed model can achieve the better segmentation performance in 1-shot and 5-shot segmentation tasks, and has better performance than the current mainstream segmentation models.