Challenge
We collaborated with Accenture to create many pilots and AI-powered platforms for the retail sector. They came to us expressly to explore if it was possible to create a system that could automatically detect shoplifting events and send an alarm to a distant screen. They also want a system that identifies empty spaces on supermarket shelves and provides alerts to workers to restock the products.
Solution
The partnership began with a definition of the most typical ways criminals steal merchandise from supermarkets. Based on it, we began working on the databases that would be used to train and evaluate the models and applications.
The approach presented was based on computer vision and deep learning techniques. The solution’s core consists in employing deep learning models to extract information about objects and important areas on the person’s body from video cameras placed along the shelves. We then used this data from tracking methods and traditional computer vision algorithms to extract temporal information from the items. Finally, all this spatial and temporal data is evaluated by hand-crafted algorithms to detect suspect behavior. The program was placed on an AWS cloud service and made available via an API for consumption by the front-end application, which displayed the camera image and the results of the algorithms.
In addition, to localize those empty spots, we construct an object identification model based on a cutting-edge neural network. We also created the database used to train the model in partnership with Accenture.
Technologies / Tools
