Edge computing optimization method. Analyzed task: crowd counting

摘要

The exponential increase in the amount of data that IoT systems generate poses serious problems to cloud and data centers, such as data management and processing scalability, power overheads, network traffic management and security. Embedded systems provided with data processing may partially mitigate such problems. Their employment is very convenient in low latency applications, in reducing traffic flows towards the cloud, in minimizing storage and processing requirements, and in providing an effective solution to several security issues. Moreover, even if with limitations, such devices allow the implementation of artificial intelligence algorithms opening to their widespread use with and without cloud integration. This work focused on the neural networks assessment on an embedded system by following a specific approach, based on edge limitations and both computational and performance parameters. The nets purpose was to detect crowded and uncrowded states. Some of these networks were then ported on an embedded platform, the STM32F767ZIT6U Nucleo board, and the trade-off between classification and computational performance was thoroughly addressed. This study shows the ability of embedded systems to run complex artificial intelligence neural networks with limited classification performance reduction.