In the near future, autonomous cars will continue to make mistakes that require human supervision to correct. To allow enough time for a human to take over, it is necessary to predict that a failure is about to occur. One approach is to try and detect input that is different from the distribution of the training set. For such input, the probability of failure is significantly higher as the car has not been trained to deal with it.
A promising out-of-distribution (OOD) detection approach are autoencoders. In this thesis, OOD detection based on autoencoders is investigated for the complex domain of urban driving. A novel task-related OOD approach based on semantic segmentation is introduced and evaluated on two large-scale driving data sets.