Authors:
Piero Herrera-Toranzo
;
Juan Castro-Rivera
and
Willy Ugarte
Affiliation:
Universidad Peruana de Ciencias Aplicadas, Lima, Peru
Keyword(s):
Stock Management, Object Detection, Computer Vision, Product Recognition, YOLOv5, Products Status.
Abstract:
Supermarkets generally do not have an efficient supervisory mechanism for inventory and warehouse management that stockists can use in their day-to-day activities. Our goal is to develop an application based on computer vision models, for the detection, counting and verification of the status of bottled and canned products. Comparisons were made between the different models for the detection of objects through an image, under the verification of parameters, performance and metrics, in order to obtain the best models. Once the YOLOv5 object detection model was chosen, training began with a dataset of own images containing products in good and bad condition in order to identify if they are damaged. Finally, the trained model was coupled to the development of the application. This application allows the user to check which products are in a loaded or taken image, as well as their quantity and status. Additionally, to facilitate the registration tasks of the storekeepers, the application
allows keeping a daily record of said products. The mAP@0.5 obtained by our model was 93.09%, while the mAP@0.5:0.95 was 89.04%. Therefore, given the results, this model can perform the task of detecting the status of proposed bottled and canned products.
(More)