Learning driving situations and behavior models from data

For an Advanced Driver Assistance System recognizing the driving situation of other vehicles is a crucial prerequisite to anticipate their behavior and plan own maneuvers accordingly. Current methods for situation recognition usually rely on an expert for defining the considered driving situations manually while solely the parameters of the corresponding behavior models are learned from observations. Unfortunately, the performance of this approach is highly dependent on the skills of the expert. Furthermore, the data for training needs to be manually labeled to define when a certain type of situation is present, which can be very time-consuming and may introduce unwanted bias. In order to circumvent these problems, we propose to learn types of situations and behavior models from data simultaneously. The goal is to identify the set of driving situations for which the corresponding behavior models achieve the best fit to given observations. As both the assignment of observations to driving situations and the model parameters are unknown, an alternating, iterative algorithm minimizing the model error is employed. We show that the algorithm accomplishes to identify reasonable driving situations and that it can be successfully applied for behavior prediction when situation labels are missing.