Effiziente Schätzung des Ego-Fahrstreifens auf RGB-Sequenzen für Mikromobilitätssysteme

In this work, we present an efficient method for ego lane detection for micro-mobility systems as electric bicycles, scooters, or tricycles using RGB sequences from a driver’s perspective. We combine a gradient-based line detector with color-based segmentation to robustly find an approximation for the ego lane borders in various traffic environments. With given geometrical conditions of the scene and temporal inference, the approximation is improved, especially in difficult cases as when driving a curve. A key task thereby is the dynamic estimation of the vanishing point using optical flow vectors between two consecutive frames. Tests on over 2000 images taken with two different recording setups and different sampling rates show that the method reliably finds the borders of the ego lane in most of the samples and approximates the ego lane in a suitable way in curves. This is confirmed with a quantitative evaluation, determining an IoU of 75.28 %. A performance of 12 fps on a Raspberry Pi 3 furthermore shows the suitability of our method for micromobility systems with low-cost and low-power hardware.