VEpred: end-to-end map-free trajectory prediction via geometry aware visual queries

Abstract Accurate and robust trajectory prediction is a fundamental prerequisite for safe decision making and motion planning in autonomous driving systems. However, most existing approaches rely heavily on known high-precision localization and explicit traffic topology maps while modeling 3D detection, object association, and trajectory prediction as separate modules. Such designs deviate from the emerging trend in end-to-end autonomous driving. To address this issue, this study proposes VEpred, an end-to-end, purely vision-based trajectory-prediction framework that directly generates future trajectories from raw sensor imagery. VEpred integrates 3D object detection, multi-object tracking, and trajectory forecasting into a unified, differentiable architecture, achieving a closed-loop representation that jointly causes perception and temporal motion within a single network, rather than through task-wise concatenation. To mitigate the potential loss of fine-grained visual cues in deep feature extraction, a visual trajectory enhancement module is introduced to align and fuse shallow edge/texture features with deep semantic features and trajectory queries, thereby improving the modeling of motion priors in dynamic scenes. Furthermore, to exploit implicit road constraints within the visual domain, an auxiliary lane-segmentation branch is incorporated to guide feature generation, enabling implicit reconstruction and enhancement of road geometry, comparable to high-definition map-based methods. Experiments on the nuScenes benchmark demonstrated that VEpred achieved an average multi-object tracking accuracy of 0.448, a recall of 56.0%, and an average displacement error /final displacement error of 1.75/2.12 within a 4.0 s prediction horizon, outperforming existing vision-only methods and even some LiDAR-based baselines. These results confirm that embedding geometric and semantic priors into dynamic trajectory queries effectively supports real-time, map-free, and end-to-end trajectory reasoning in autonomous and cooperative driving scenarios.