Towards a Framework of Open-Source Hardware and Software Components for Real-Time and Custom-Made Human Activity Recognition based either on Inertial Sensors or on Vision Sensors

Human Activity Recognition (HAR) consists of detecting and identifying human activities throughout analysis of sensor measurements (wearable sensors, ambient sensors, optical sensors, etc). HAR is mainly structured by studies based on inertial sensors and vision sensors and rises as an important topic nowadays as finely and precisely measuring human activities is key and fundamental in several domains such as health, sport, security, etc. The ubiquity of sensors, cameras, and artificial intelligence based applications in our daily lives makes it possible to benefit from join analysis of both sensor measurement and camera image processing. However, hardware and software components that support automatic monitoring and analysis of human activities still remain unaffordable either because it is expensive to acquire such a complete commercial solution or because it is difficult to getting started with such different technologies coming from various fields. This paper presents the first version of a framework with several open-source hardware and software components that makes real-time sensor-based or vision-based HAR processing possible. It shows how to build custom-made datasets and models based on data collected from inertial sensor measurements and from camera images through a study case of four daily-life activities with two motion activities and two motion-less ones. It also shows how to perform real-time processing of inertial sensor-based HAR and vision-based HAR.