Machine-Learning Application in Surveillance and Infrastructure Radars — Small Target Detection and Identification of Vulnerable Road Users

This presentation consists of three parts. The first describes a recently proposed algorithm for detecting small targets in a known surveillance area using an autoencoder-based detection scheme. In this algorithm, an autoencoder learns to reconstruct the scene under observation without the static background and with an emphasis on small non-moving targets. The output of the autoencoder is then used to detect targets with a threshold. The results of this algorithm are shown for real measurements of a 94GHz radar with hovering drones and a person as targets. The radar is a scanning surveillance radar with a real aperture imaging capability. However, the autoencoder is applied to range profiles to reduce the computational effort. The second part of this presentation describes the results of an infrastructure project. The result of this project was a data fusion system that combines infrared and radar sensors to detect vulnerable road users, ie cyclists and pedestrians. The aim of this system is to warn approaching cars when people are walking towards the road at junctions or bus stops, thereby reducing the number of accidents. Details of the CNN-based radar classification system are given in this presentation together with some basic elements of the necessary micro-Doppler data processing. This processing is used to simulate data from public motion capture data to support the training process of the neural network. Furthermore, explainable AI tools such as Shap are used to find the critical features used for classification. The last part describes some potential use cases of the recently proposed federated learning scheme for distributed radar systems. Federated learning has been introduced for edge devices, such as phones and tablets, to learn from other people’s data without seeing their data. This can also be applied to radar networks to learn from other radars without sharing the data. Reasons for avoiding data sharing could be privacy issues or simply the data load and limited communication capabilities. The basic principles of this scheme are presented with some example applications.