Human Activity Recognition (HAR) is defined as the detection and interpretation of human activity. In the context of Ambient Assisted Living (AAL), for example, this can mean capturing the actions of individuals to enable them to live as long as possible in their desired environment in a self-determined manner. Other use cases include optimizing training sessions in sports science, or supporting individuals in their craft training. The HAR must fulfill several fundamental aspects that go beyond pure technical performance: It must be as low-maintenance as possible, easy to install, inexpensive to maintain, and simple to operate.
For the analysis of Human Activity Recognition (HAR), artificial intelligence (AI) is increasingly used in addition to classical signal processing methods. At the beginning of an AI development for HAR, due to the diversity of the design of a HAR system, there is always the trade-off between the AI algorithms to be applied and the partitioning of the AI: parts of it can be executed as embedded AI on the sensor node itself, other parts run centrally on a local AI computer or in the cloud. This creates a hierarchical and modular AI architecture that aims to achieve good performance while being resource-efficient.
In order to design a Human Activity Recognition efficiently and effectively, it is necessary to adequately capture the human activity by means of measurement. The central point here is the realistic recording of actions through suitable parameters with the highest possible resolution: whether it is the analysis of the human pose in 3D-space by means of computer vision, or the measurement of the accelerations of the individual human extremities with wearable devices. Indirect sensing can also be applied here, where the objects that humans interact with during a given process are equipped with sensing devices. The goal is thus to design sensory hardware systems and entire sensor networks that provide these capabilities.