Generating and Using Synthetic Data for Machine Learning in Personnel Security Screening Scenarios

State-of-the-art personnel security screening based on microwave imaging technology heavily relies on threat detection models created by using supervised machine learning techniques. This usually calls for excessive efforts to build up huge databases containing labelled sample data of people being screened. In the case of walk-through screening devices, threat detection may also incorporate a posture tracking mechanism of the person being screened. This normally calls for the creation of a dedicated posture database through the tedious and costly task of manually labelling postures in volumetric microwave images. However, it is demonstrated that by employing electromagnetic simulation, a comparable posture tracking algorithm can be created while using synthetic data. Thus, the development time and costs can be reduced while maintaining a high threat detection performance. In addition, synthetic data can improve posture detection accuracy or add support for novel postures.