Video-Based Gender Profiling on Challenging Camera Viewpoint for Restaurant Data Analytics
Jefferson James U. Keh, Meygen Cruz, Ramiel Deticio, Carl Vincent Tan, John Anthony C. Jose, Elmer P. Dadios, Alexis M. Fillone · 2020
The customer gender composition is an important consideration for certain business decisions. This study highlights the feasibility of implementing a vision-based gender classification feature on top of a pre-existing people counting system that makes use of a CCTV camera in a retail establishment. The main problem lies in doing so using the fixed angled viewpoint of the restaurant's camera, which represents a practical use case as CCTV systems are often geared for security purposes instead of data analytics. Thus, these existing camera configurations are not ideal for existing computer vision models. A two-step neural networks-based approach, namely head detection via YOLOv2 and gender classification via Inception V3, is used to integrate this feature into the pre-existing system. The researchers partnered with a restaurant in a business district to obtain footage and test the system. The proponents were able to attain the target accuracy of 80%, albeit the feature's speed and accuracy can be further improved through the exploration of other object classification algorithms and through the use of downsizing and down sampling.