A Mapless Trajectory Prediction Model with Enhanced Temporal Modeling

The prediction of agents' future trajectory is a crucial task in supporting advanced driver-assistance systems (ADAS) and plays a vital role in ensuring safe decisions for autonomous driving (AD). Currently, prevailing trajectory prediction methods heavily rely on high-definition maps (HD maps) as a source of prior knowledge. While HD maps enhance the accuracy of trajectory prediction by providing information about the surrounding environment, their widespread use is limited due to their high cost and legal restrictions. Furthermore, due to object occlusion, limited field of view, and other factors, the historical trajectory of the target agent is often incomplete This limitation significantly reduces the accuracy of trajectory prediction. Therefore, this paper proposes ETSA-Pred, a mapless trajectory prediction model that incorporates enhanced temporal modeling and spatial self-attention. The novel enhanced temporal modeling is based on neural controlled differential equations (NCDEs) and vanilla temporal self-attention mechanism to learn temporal features from incomplete historical trajectories. By combining this enhanced temporal modeling and spatial self-attention, our model fully captures spatio-temporal features within the scene. Experiments results on the nuScenes dataset demonstrate the superior performance of our model compared to existing mapless trajectory prediction models. Comprehensive ablation studies further confirm the effectiveness of each proposed module in our model.