Estimating the Uniqueness of Test Scenarios derived from Recorded Real-World-Driving-Data using Autoencoders

Advanced Driver Assistant Systems (ADAS) use a multitude of input signals for tasks like trajectory planning and control of vehicle dynamics provided by a large variety of information sources such as sensors and digital maps. To assure the feature's valid behavior all realistically possible environmental situations have to be tested. The test scenarios used for simulation can be derived from real-worlddriving-data. However, the significance of derived scenarios is weakened by repetitive similar situations within the driving data, which increase the test efforts without providing new insights regarding the test of the ADAS. In this contribution, an automated selection algorithm for test scenarios based on relevant environmental parameters is presented. Starting with a randomly selected initial testset, the machine-learning concept of autoencoders is utilized to recognize novel scenarios within the data pool, which are iteratively added to the initial testset. Furthermore, the key parameters for the autoencoder's performance are shown in depths. The approach is fully automated, so that the identified novel scenarios within an entire testset are automatically combined to a reduced testset of unique relevant scenarios. The achieved testset reduction and thereby the saving potential in simulation time is demonstrated on a dataset including several thousand test kilometers.