Freespace and Optical Flow Fusion for Generic Contours Accumulation

Autonomous driving heavily relies on strong safety cross-checks. When it comes to safety, combining multiple measurement paths to decompose the risk is an industry standard. This work proposes a novel fusion solution of classical video-based signals and modern Deep Neural Networks outputs for stabilizing Freespace Contours. We propose the usage of a well-established measurement in autonomous driving, Optical Flow, in a complementary usage with Deep Neural Network-based Freespace contours to enable a temporally stable representation of very complex autonomous driving scenarios. Enhancing plausibility checks of Freespace contours with Optical Flow increases safety and robustness of the output in difficult corner cases. By introducing our method, Freespace contour of dynamic objects are stabilized over time.