Comparing autoencoder-based approaches for anomaly detection in highway driving scenario images

Abstract Autoencoder-based anomaly detection approaches can be used for precluding scope compliance failures of the automotive perception. However, the applicability of these approaches for the automotive domain should be thoroughly investigated. We study the capability of two autoencoder-based approaches using reconstruction errors and bottleneck-values for detecting semantic anomalies in automotive images. As a use-case, we consider a specific highway driving scenario identifying if there are any vehicles in the field of view of a front-looking camera. We conduct a series of experiments with two simulated driving scenario datasets and measure anomaly detection performance for different cases. We systematically test different autoencoders and training parameters, as well as the influence of image colors. We show that the autoencoder-based approaches demonstrate promising results for detecting semantic anomalies in highway driving scenario images in some cases. However, we also observe the variability of anomaly detection performance between different experiments. The autoencoder-based approaches are capable of detecting semantic anomalies in highway driving scenario images to some extent. However, further research with other use-cases and real datasets is needed before they can be safely applied in the automotive domain.