Abstract:

Existing trajectory prediction methods often overlook the interaction between vehicles and the map， resulting in trajectory predictions that do not conform to road topologies. To address this issue， this paper proposes a target-anchor-driven multimodal trajectory prediction method （TMTP） that couples the vehicle's motion trend. The proposed model efficiently incorporates prior knowledge of traffic scenarios into the algorithm through a graph model， allowing for precise description of heterogeneous interactions within traffic scenes. The model thoroughly considers the interaction between the vehicle's historical trajectories in the dynamic scene graph， the future trajectories of the ego vehicle， and the topological information of the vectorized map in the static scene graph. By utilizing an attention network， the model aggregates features from different nodes， achieving enhanced local-global feature fusion. Furthermore， TMTP represents driving intentions as target anchors， simplifying the complexity of the intention space. The proposed method was evaluated on the large-scale Argoverse motion forecasting benchmark. The results demonstrate thatthe model introduced in this paper outperforms the official benchmark model by 56.2% and 56.6% in metrics \mathrm{m}\mathrm{i}\mathrm{n}\mathrm{F}\mathrm{D}{\mathrm{E}}_{\mathrm{1}} and \mathrm{m}\mathrm{i}\mathrm{n}\mathrm{F}\mathrm{D}{\mathrm{E}}_{\mathrm{6}}， respectively， exhibiting an exemplary capability in accomplishing the task of trajectory prediction.

Key words: vehicle engineering, trajectory prediction, target anchor, graph neural networks