Segmentation based Non-learning Product Detection for Product Recognition on Store Shelves

The arrangement of products on store shelves can refer to commercial contracts, sale achievement, and customer satisfaction. At present, clerks check the arrangement manually, which spends time, costs human resource significantly, and can disturb shopping customers. Although automatic methods via computer vision (often incorporating machine learning) can solve the issue, the existing methods need single product template images for detection, facing the difficult collection of master images and the frequent upgrade of products. In this paper, we propose to detect products on store shelves by segmenting the shelves horizontally and vertically without template images and machine learning. The horizontal segmentation is based on clapboard detection via casting lateral gradient votes. The vertical segmentation contains the linear region of interest (ROI) optimization and shadow detection by longitudinal gradient grouping. In experiments, we compare our method with the only existing non-template method, and our method outperforms the existing method.