G2LTraj: A Global-to-Local Generation Approach for Trajectory Prediction

TL;DR

G2LTraj tackles trajectory prediction by bridging simultaneous and recursive approaches through a global-to-local generation framework. It first generates globally distributed key steps with spatial constraints, then recursively fills intermediate steps within local sections while incorporating agent features for stronger temporal consistency, and finally selects an optimal granularity via a learnable confidence mechanism. The method yields consistent improvements across ETH, UCY, and nuScenes over multiple baselines and exhibits robust ablations, with insights into latency, hyper-parameter sensitivity, and the benefit of adaptive granularity for diverse motion patterns. This approach advances kinematically feasible, multi-step trajectory forecasting with practical applicability for autonomous driving and real-time systems.

Abstract

Predicting future trajectories of traffic agents accurately holds substantial importance in various applications such as autonomous driving. Previous methods commonly infer all future steps of an agent either recursively or simultaneously. However, the recursive strategy suffers from the accumulated error, while the simultaneous strategy overlooks the constraints among future steps, resulting in kinematically infeasible predictions. To address these issues, in this paper, we propose G2LTraj, a plug-and-play global-to-local generation approach for trajectory prediction. Specifically, we generate a series of global key steps that uniformly cover the entire future time range. Subsequently, the local intermediate steps between the adjacent key steps are recursively filled in. In this way, we prevent the accumulated error from propagating beyond the adjacent key steps. Moreover, to boost the kinematical feasibility, we not only introduce the spatial constraints among key steps but also strengthen the temporal constraints among the intermediate steps. Finally, to ensure the optimal granularity of key steps, we design a selectable granularity strategy that caters to each predicted trajectory. Our G2LTraj significantly improves the performance of seven existing trajectory predictors across the ETH, UCY and nuScenes datasets. Experimental results demonstrate its effectiveness. Code will be available at https://github.com/Zhanwei-Z/G2LTraj.

Figures (8)

• Figure 1: (a) State-of-the-art simultaneous approaches like GraphTERN bae2023set do not model the constraints among future steps well, resulting in kinematically infeasible predictions (i.e., the fluctuation in the predicted trajectories), as evaluated on the UCY lerner2007crowds dataset. (b) Our G2LTraj incorporates spatial-temporal constraints to generate consistent trajectories. (c) Overview of our G2LTraj for trajectory prediction in a global-to-local generation process. For instance, initially, we simultaneously generate the global key steps $t_1$, $t_9$ and so on. Subsequently, $t_5$ is generated by utilizing $t_1$ and $t_9$. This iterative process continues until all future steps are generated.
• Figure 2: The overview framework of our G2LTraj. We introduce a global-to-local generation approach for trajectory prediction that attenuates the accumulated error and proposes constraints among future steps. In this illustration, we exemplify the global-to-local generation process at a granularity of 8. we simultaneously generate the globakey steps $t_1$, $t_9$ and so on. Subsequently, $t_5$ is generated by utilizing $t_1$ and $t_9$. This local recursion process continues until all future steps are generated.
• Figure 3: Merely relying on $t_5$ and $t_9$ limits the model's access to historical kinematic information, possibly generating plausible but kinematically infeasible $t_7$, as shown in the red oval. By incorporating historical kinematics clues, the prediction range can be reduced to the kinematically feasible range, as shown in the black oval.
• Figure 4: The accumulated error comparison at each future step. Accumulated Error (5) and (10) are the average of pointwise L2 distance between each future step and ground truth over 5 and 10 most likely predictions respectively.
• Figure 5: Qualitative results on the UCY validation set. The black ovals denote a failure case.