A Method for Designing an Embedded Human Activity Recognition System for a Kitchen Use Case Based on Machine Learning

Abstract

Human activity recognition enables technical systems to analyse human behaviour in various settings. For example, it can be directly used to support the user in elder care, healthcare or training environments. Nevertheless, human activities are often times highly variable and therefore pose a challenge for any technical system to correctly classify and, even more importantly, generate a feedback that is valuable to the user. In this paper the process for designing a system that uses machine learning on the sensor node itself is presented in order to improve human activity recognition within a sensor network. Each sensor node of the network consists of a Bluetooth capable system on module and an accelerometer. The acceleration data is used to distinguish between several slicing techniques of different vegetables with the aim to help the network to distinguish the different dishes cooked with those vegetables. Various steps were taken to find the best possible machine learning model and sensor configuration to infer the cut vegetable on the sensor hardware, which is based on a standard microcontroller and therefore poses a challenge with its limited memory. Overall, the system is able to correctly infer the correct class most of the times while enabling a sufficient battery run time. Within this paper these steps and tests for the design and implementation of the embedded machine learning algorithm is described and its capability for activity recognition evaluated