Research on Map Update Method Based on Online Learning Optimization with multi-modal Data

In the field of autonomous driving, the accuracy of High-Definition map (HD map) is crucial. However, survey accuracy is reduced by the complex road structures and frequent terrain changes in open-pit mines. In addition, frequent manual updates are required for maps, leading to significant cost issues. To address these issues, a map update method is proposed based on online learning optimization using multi-modal data. The method first constructs the topological structure of initial map reference lines and road boundaries, and discretizes the sequence of points based on map structure, connectivity, and real-time perception. Then, the Frenet coordinate system is used to obtain the vehicle's real-time path deviation relative to the reference line. These multi-modal data are integrated based on the map sequence to create a dataset. Next, a learning network based on the Transformer architecture is proposed. This network is trained using the self-constructed dataset to output updated map reference lines and boundary lines. Finally, the effectiveness of the algorithm is tested on a segment of road with changes in an actual mining truck operating scenario. Results show that the proposed method can update road curves in real time. The maximum update error for the road centerline is 0.093m and the average update error is 0.049m. For the boundary line, the maximum update error is 0.095m and the average update error is 0.058m.