关键词: Transformer, 单目视觉, 存量检测, 深度学习, 聚类算法

Abstract: To address the visual recognition difficulties caused by non-structured item placement, occlusion, specular reflection, and frost formation inside household freezers and refrigerators, a monocular depth estimation method based on Vision Transformer is proposed for inventory status detection and low-inventory region localization in household freezers and refrigerators. The method uses Vision Transformer to extract global features and employs a convolutional decoder for multi-scale feature fusion, thereby generating high-resolution depth maps. On this basis, the predicted depth map is compared with the empty-shelf reference depth map, and depth-difference analysis and K-means clustering are combined to identify low-inventory regions and determine inventory status. The model is pre-trained on the NYU Depth V2 dataset and fine-tuned on a self-built household freezer and refrigerator scene dataset. Experimental results show that the proposed method achieves an accuracy of 96.26% in the binary inventory status classification task, outperforming object detection methods such as YOLO v3, YOLO v5, and RT-DETR. It also demonstrates good robustness under complex occlusion and lighting conditions. The results indicate that the proposed method provides an efficient and low-cost technical approach for intelligent inventory monitoring in household refrigerators and freezers.

Key words: Transformer, Monocular vision, Inventory monitoring, Deep learning, Clustering algorithms