Testing Scenario Identification for Automated Vehicles Based on Deep Unsupervised Learning

Naturalistic driving data (NDD) are valuable for testing autonomous systems under various conditions. Automatically identifying scenes from high-dimensional and unlabeled NDD remains a challenging task. This paper presents novel approach automatically test scenarios through deep unsupervised learning. Firstly, US DAS2 leveraged, the selection of variables representing vehicle state surrounding environment is conducted to formulate segmentation criterion. The isolation forest (IF) algorithm then employed segment data, yielding two distinct types datasets: typical extreme scenarios. Secondly, one-dimensional residual convolutional autoencoder (1D-RCAE) developed extract scenario features datasets. Compared four other autoencoders, 1D-RCAE can effectively crucial information with optimal feature extraction capability. Next, considering varying importance different features, an entropy (IE)-optimized K-means cluster extracted using 1D-RCAE. Finally, statistical analysis performed on parameters each explore their distribution characteristics within class, identified along five proposed framework, combining IF, 1D-RCAE, IE-improved algorithms, identify NDD. These be applied performance systems, enriching library automated