The smart home market has experienced considerable growth in recent years. Nevertheless, the classic light switch remains a central control element in private households. Motion detectors are often impractical in frequently used living spaces, as they cannot distinguish between different usage scenarios – such as watching TV and eating. Touchless gesture control offers a promising alternative. However, optical sensors such as cameras cannot meet the requirements for data protection and robustness in varying lighting conditions. The aim of this work is to develop and evaluate a system for the reliable detection and direction classification of arm gestures using two radar sensors, which are not affected by the disadvantages of optical sensors. The double detection by two sensors enables a more robust interpretation of the scene from different perspectives. A continuous upward movement of the arm is defined as a specific gesture. The work begins with an introduction to the theoretical foundations of radar technology and relevant concepts of machine learning, in particular neural networks. The central component of the system is a neural network based on a variational autoencoder, which performs several tasks simultaneously. It reduces the noise in the radar data, extracts relevant movement information and arranges it in an intermediate representation in such a way that different directions become distinguishable. On this basis, the gesture direction is classified into eight discrete directions with an angular spacing of 45∘ and gestures are distinguished from non-gestures. The evaluation of the developed neural network is carried out retrospectively on a computer using recorded radar data. The work examines how reliably the system can classify the direction of arm gestures and distinguish them from non-gestures. This results in a very high classification accuracy of 98.8 % in known and 96.3 % in unknown environments, in each case for clearly delimited gestures within a defined area. In addition, the variational autoencoder effectively reduces noise in the radar signals and preserves the relevant movement characteristics of the gestures. To evaluate the system’s practical applicability, additional realistic test scenarios with natural motion sequences are carried out. With the help of additional decision logic, the number of misclassifications in these scenarios can be significantly reduced. The combination of two radar sensors enables a more robust classification of direction and improves the accuracy of occluded movements.
Detektion und Richtungsbestimmung von Armgesten mittels FMCW-Radarsensoren und künstlicher Intelligenz
Stellnberger, S. (Author). 2025
Student thesis: Master's Thesis