MA 1744: Evaluation of Neural Networks in Radar Signal Processing for Human Activity Recognition
Smart home applications are increasingly finding their way into the everyday lives of many people. Audio and camera systems have already established voice and gesture control in some households. However, the audiovisual sensor technology is a relatively strong intrusion into people’s privacy. Radar sensors offer a decisive advantage in this respect. Since the data do not have any immediately recognizable personal reference. For both private and public buildings, radar sensors therefore offer many attractive possibilities for performing tasks, which involve human behavior, even with high data protection requirements. They can be used, for example, to detect people in the building, their general movements and even to measure vital parameters.
In state-of-the-art radar signal processing, several consecutive Fourier transforms are applied to the radar raw data to identify the features of the targets. Based on this pre-processed data, human activities are classified using machine learning techniques by feeding the data into a classifier network (neural networks / deep learning). This enables the recognition of gestures and specific motions of humans. However, another option is to directly replace the pre-processing steps with an AI network. This would make it possible to carry out the entire chain from raw data to classification in an AI network.
The task of the master´s thesis would therefore initially be to map the radar processing steps in a neural network. In order to generate sufficient radar training data, we work with artificially generated data using ray-tracing software. For this purpose, a realistic 3D model of a room with moving people has to be created (e.g. in Blender software). Furthermore, a neural network for replacing the processing steps is trained and evaluated. Moreover, a neural network for human activity classification has to be designed in order to compare the classification task with and without possible modifications in the input data due to the developed network. Finally, both networks are combined into one and the performance is compared. For validating the developed network, test measurements are carried out and compared with the state of the art.
Edited by Jianyang Wu.