Crowdsourced HD Map Patches Based on Road Model Inference and Graph-Based SLAM

In this work, we propose a method to create crowdsourced HD map patches based on road model inference and graph-based SLAM. As input data, we use higher level features such as lane marking and traffic sign observations that, in near future, will also be available from a newly developed floating car data collection solution. We process the data in four steps: (i) we apply smoothing based on GPS and odometry readings, (ii) align the resulting traces based on common observations of 3D point landmarks, (iii) find the lane marking configuration sequence along the road by discrete optimization and (iv) use that sequence to associate lane marking observations with lane marking entities to obtain their geometry. The results are evaluated by comparison with a ground truth HD map. Based on a total of 29 km of learned road segments, the average absolute distance between learned lane marking positions and those in the map is 0.306 m.