Section: New Results

Brick & Mortar Cookies

Participants : Julien Badie, Manikandan Bakthavatchalam, Anais Ducoffe.

BMC is a software platform that was introduced in 2016. The goal was to create a light and easy to deploy platform in the context of gathering Business Intelligence data (BI) to help store managers monitor client trajectories and most interesting products and alleys inside a supermarket. BMC was finalized this year and had been extensively tested in real conditions with success.

The platform is divided into several modules :

• an image processing module that handles retrieving images from a camera network followed by people detection using a deep learning algorithm (SSD) and tracking. The results are stored in a MySQL database;

• a trajectory analysis module that computes statistics for BI. These statistics are computed based on the trajectories of the first module and give information regarding each alley such as the average number of customers per hour, the average time spent by the customer in the alley or the average number of stops in front of each type of product. All these data can be visualized via a web interface and are refreshed periodically;

• an automatic installation and deployment module. This module is a set of Python scripts that, given a computer with the correct hardware and OS requirements will automatically install the BMC platform and all its dependencies (OpenCV, cuda, caffe, ...). In a second time, users can enter the list of cameras they want to process with a limited set of parameters (working days and hours) and the deployment script will prepare the platform and create the CRON command line that will run BMC automatically.

An additional module was created but not included in the release : the evaluation module. It was used to find the best balance between fast processing and precise results by evaluating different parameters at the people detection level such as which model to use and which image resolution to process. After several experiments, it was decided to use a processing resolution of 480x272 to allow us to process 16 cameras simultaneously with two Nvidia GeForce GTX 980 GPU at 3-4 FPS. This framerate is sufficient for tracking in the context of a supermarket as people tends to move slowly and stops a lot. The platform was tested in a supermarket in Nice during several days and showed satisfactory results.

BMC is registered at the APP under the name BMC_1. It is intended to be used in the FUI project StoreConnect.