von Herrn Alexander Scheible, M. Sc.
Kurzfassung:
Model-based object tracking relies on stochastic models that describe how objects evolve over time and which measurements are expected measurements from a sensor. Together with Bayes theorem, these filter models are used to update prior knowledge with new measurements. This works great if the models are accurate, however, the performance breaks down if the models do not fit the data. To address this, I suggest to work from two sides: First, increasing the performance of the algorithms themselves, and second, monitoring the system to recognize when it is not working as intended.
My talk starts with the first part. I present a method for unsupervised learning of filter models directly from measurements, using the multi-object measurement likelihood as loss.
The second part gives an overview of monitoring methods. I conclude with open questions and possible future research directions.
Zu diesem Vortrag wird recht herzlich eingeladen!