Computer Vision with Cognitive Learning to Improve the Decision-Making During the Sales Process in Physical Stores

Abstract

In a world where the information is obtained faster than ever seem, new methods to process that high volume of data are being developed frequently. This is more notorious in a virtual ambient where the data is generated in a manner that is faster and easier to analyze than in the real world. This is very evident in the retail field, where virtual stores have easy access to all the advertisement a user visited and simple to obtain user profile, on the other hand physical stores are limited to basically create a register in a database when there is a purchase. In an attempt to improve the retailers experience from physical stores to manage their business this document has the objective to develop a computational tool that will analyze the people flux going in the establishment, trying to inform the retailer the amount of people and their gender to help the sales process in physical stores. To this end, computational vision methods and algorithms were raised, which after selection, theoretical conception and tool’s implementation it was tested with benchmarks to operate locally and in real time by accessing the cameras installed strategically in a real scenario. Two scenarios were tested: static ambient light and dynamic light. Two tests were conducted: YOLOv2 against background subtraction-based counter; gender classification using full body features. Even though the results were not as positive as needed for commercial use, the tool demonstrated potential and space for improvements.