Anomaly detection for automotive visual signal transition estimation

We developed an end-to-end pipeline for brake light transition detection based on cognitive theories of anomaly-detection and model based systems engineering principles. Inspired by cognitive theory, we decompose the visual input stream into the submodalities color, shape, intensity and motion which is closely coupled with a graphical model. A memory module that contains priors is populated by exploitation of knowledge from specifications and simulation. High-fidelity 3D-Simulations have been created to populate motion memory, whereas low-fidelity 3D-projections have been used to create geometric priors. We captured real world sequences containing RGB images, GPS- and inertial measurement-data and annotated them with boundingboxes for cars and their tail and brakelight states, which are used to evaluate the proposed pipeline.