Scenario Detection in Unlabeled Real Driving Data with a Rule-Based State Machine Supported by a Recurrent Neural Network
Francesco Montanari, Haoyu Ren, Anatoli Djanatliev
Abstract
An arising idea in the automotive sector is to extract and collect scenarios from real driving data and use them as test cases for the validation of automated driving functions. In this paper, we use a rule-based state machine to label the data for the training of a recurrent neural network (RNN) and combine both the state machine and the RNN for detecting driving scenarios. The state machine shows precise results and the idea of training the RNN on the resulted samples from the state machine shows promising results. A statistical comparison of the proposed methods shows that the state machine should be used if possible, however, if the signals needed for the state machine are not available the RNN can be used to support it.