Ghost target classification using scene models in radar

Sammanfattning: In surveillance contexts, radars can be used to monitor an area, detecting and tracking moving objects inside it. Monitored areas in urban environments often contain many surfaces that reflect radar waves, which can have the undesired consequence of a single object producing multiple tracks due to multipath propagation effects. This thesis considers a method of identifying if a track is produced by a real object, or if it stems from multipath effects. The proposed method works by creating a machine-learning-based classifier and modelling the monitored scene over time. Tracks are assigned features based on their characteristics and the state of the scene model in regards to their position. These features are then used as inputs to the classifier model to produce the classification. We propose four machine learning-based classifier models, with two different sets of structures and features used. The classifier models are compared to a naive classifier model for reference. The proposed models all outperform the naive classifier, although some of them are biased. As for the usefulness of the scene model, the results are mixed but show promise. We believe that the scene model can improve classification performance further with more and better data.