NavPathNet: Onboard Trajectory Prediction for Two-Wheelers using Navigation Maps

This work proposes NavPathNet, a graph neural network based trajectory prediction system capable of accurately predicting the trajectory of a two-wheeler with onboard sensors for up to 6 seconds using the vehicle state and only navigation map information. Possible paths are generated from the local road network in a regular navigation map using Bezier curves and a multimodal prediction captures the different possibilities of the road network. A kinematic model integrated into the network allows to give guarantees on physical realism and on motion constraints, essential for safety-relevant systems. The proposed system was trained and evaluated on an in-house e-bike data set and was able to reach a top 1 final displacement error of 2 m for four seconds and 3.8 m for six seconds of prediction time, significantly outperforming other baselines. The improvement in prediction quality compared to purely physical models opens up new possibilities for driver assistance systems in connected vehicles.