Development of a POS System with Computer Vision for Automated Retail Checkout

Abstract

A Point-of-Sale (POS) is a computerized system of hardware and software utilized by businesses to complete sales transactions. In conventional POS setups, cashiers manually scan individual product barcodes, before processing the totals. This manual procedure is laborious and often leads to long queues and waiting times, especially during peak hours, ultimately affecting customer experience and retention. This work seeks to automate the product scanning procedure with a computer vision approach, thereby expediting the sales process. An efficient YOLOv4 object detection model was trained on a custom dataset of common products found in Malaysian retail stores. 550 images were initially acquired and split 80:20 into training and validation groups; further augmentation tripled the size of the training group to 1,320 images. Training was conducted for 10,000 epochs, at 0.0013 learning rate. During training, the model achieved 99.19% mAP, 87.42% average IoU, and a 0.40 average loss. Subsequently, the model was deployed on a low-power single-board computer running a transaction notification program. To evaluate its performance, 10 instances of shopping carts with random product combinations were processed using the system. The system autonomously identified and quantified all products through its video feed, generating itemized bills in real-time. Fixed with a 0.9 confidence threshold, the system yielded a 98% average accuracy across all object classes. On average, transactions, from product detection to delivering the itemized bill to the system administrator, were processed in just 14 seconds. This POS system holds potential for integration with unmanned stores, offering a seamless shopping experience.