Balancing between centralized vs. edge processing in IoT platforms with applicability in advanced people flow analysis

Understanding consumer behavior has always been a major concern for any business. Although in the last decades there has been a steep rise in online shopping and online marketing, a big chunk of the revenue for retailers still comes from traditional, in-store shopping. As such, analyzing and understanding the flow of customers inside supermarkets and stores can offer invaluable insights regarding their business and customer behavior to retailers. With the rise of smart IoT devices that allow live recording of video streams and even processing these streams on-edge, analyzing people flows in the hyperspace of crowd size, person categories (by sex and age), person location and time can prove to be an essential tool for decision makers. This paper presents a method for performing people flow analysis using deep learning for person recognition and profiling. The proposed method is encapsulated in a platform that can analyze the people flow on-edge or on a central server with powerful GPUs. The platform also encompasses several business intelligence graphical dashboards for presenting and analyzing the resulted time series data. Our preliminary findings show that on-edge processing, on embedded devices, is a plausible alternative to central processing with GPUs. Despite having lower performance, both in term of mean average precision and frames per second, embedded devices and the corresponding algorithms still manage to achieve reasonable results, while being cheaper and significantly more power efficient. Having this said, the most reasonable conclusion is to combine the two solutions mentioned above into a hybrid approach.