CH-016
Estimation of Product Amount on Store Shelves Using Image Change Classification Based on Background Subtraction and CNN

○比嘉恭太・岩元浩太（NEC）

This paper proposes a method to estimate product amount on store shelves from a video captured from a fixed camera attached on the ceiling. First, the proposed method detects change regions in the image using background subtraction based on statistical information of pixels. Then, the detected image change regions are classified into four classes representing the actual change in product amount such as "product taken (decrease)" and "product replenished (increase)" by using a convolutional neural network. Finally, the display condition (presence or absence) of products is accurately updated using classification results, and then product amount on shelves is computed. Experimental results using videos captured in a real store show that the proposed method achieves 89.2% accuracy when tolerance is approximately one product. With the high accuracy, store clerks can replenish products when stockout occurs, enabling the reduction of sales opportunity loss in retail stores.