SparScene: Efficient Traffic Scene Representation via Sparse Graph Learning for Large-Scale Trajectory Generation

TL;DR

Sparsity and topology priors are leveraged to tackle scalable multi-agent trajectory generation in complex traffic. SparScene builds a symmetric, lane-topology guided graph and a three-stage encoder (TiL, L2A, A2A) to propagate information efficiently, paired with a lightweight decoder for multimodal predictions. Empirical results on WOMD show competitive accuracy with millisecond-level inference and strong regional-scale scalability, outperforming dense-attention baselines in efficiency. This approach offers a practical path to real-time, large-scale traffic scene generation for autonomous driving and simulation systems.

Abstract

Multi-agent trajectory generation is a core problem for autonomous driving and intelligent transportation systems. However, efficiently modeling the dynamic interactions between numerous road users and infrastructures in complex scenes remains an open problem. Existing methods typically employ distance-based or fully connected dense graph structures to capture interaction information, which not only introduces a large number of redundant edges but also requires complex and heavily parameterized networks for encoding, thereby resulting in low training and inference efficiency, limiting scalability to large and complex traffic scenes. To overcome the limitations of existing methods, we propose SparScene, a sparse graph learning framework designed for efficient and scalable traffic scene representation. Instead of relying on distance thresholds, SparScene leverages the lane graph topology to construct structure-aware sparse connections between agents and lanes, enabling efficient yet informative scene graph representation. SparScene adopts a lightweight graph encoder that efficiently aggregates agent-map and agent-agent interactions, yielding compact scene representations with substantially improved efficiency and scalability. On the motion prediction benchmark of the Waymo Open Motion Dataset (WOMD), SparScene achieves competitive performance with remarkable efficiency. It generates trajectories for more than 200 agents in a scene within 5 ms and scales to more than 5,000 agents and 17,000 lanes with merely 54 ms of inference time with a GPU memory of 2.9 GB, highlighting its superior scalability for large-scale traffic scenes.

Figures (7)

• Figure 1: SparScene proposes a topology-guided interaction modeling method that selectively models topology-consistent interactions with strong behavioral relevance, resulting in sparse and interpretable interaction graphs. Upper: A local traffic scene from the perspective of the target agent. Bottom left: Distance-based interaction modeling produces dense and redundant connections and may miss behaviorally important long-range interactions. Bottom right: Topology-guided interaction modeling in SparScene yields sparse, interpretable interaction graphs by capturing topology-consistent interactions beyond distance constraints. Note: SparScene adopts a symmetric scene representation for all agents and lanes (see Fig. \ref{['fig: Symmetric_Scene']}).
• Figure 2: Scene graph initialization for SparScene. The leftmost panel shows a local traffic scene with agents and lane geometries. The second and third panels illustrate lane node extraction and lane graph construction from the HD map, respectively. The rightmost panel depicts the agent-to-lane interaction edges for the target agent.
• Figure 3: Symmetric scene representation.Left: Global scene representation with a fixed origin and orientation. Right: Symmetric scene representation, where each agent and lane is represented in its own local coordinate system.
• Figure 4: Architecture of the SparScene framework. A symmetric traffic scene is built by representing every agent and lane segment in its own local coordinate frame, ensuring a coordinate-invariant representation. Based on spatial agent–lane alignment and HD-map connectivity, initial agent-to-lane ($M_{A2L}$) and lane–to–lane ($M_{L2L}$) edges are constructed as priors. Agents and lanes are encoded with dynamics and geometric semantics, while edge attributes are embedded into a shared latent space. The encoded nodes are processed by the proposed Lane-Topology Guided Scene Encoding (Sub.Sec.\ref{['subsec: scene_enc']}), which consists of three stages: 1) Traffic-in-Lane (TiL) aggregates local traffic dynamics into lane nodes; 2) Lane-to-Agent (L2A) propagates traffic-aware lane semantics back to agents; and 3) Agent-to-Agent (A2A) builds sparse, behaviorally feasible interactions induced by lane topology. The resulting representation supports efficient and accurate multi-agent multimodal (MAMM) trajectory prediction with a lightweight decoder.
• Figure 5: Number of edges against A2L radius. Ablation of the A2L connection radius $r$ under a fixed OFF topology search strategy. The left y-axis shows the average number of edges per scene for different edge types (over 100 scenarios in the validation set), while the right y-axis reports the corresponding prediction error measured by minFDE@8s (over 41,000 scenarios in the validation set). Increasing $r$ leads to a rapid growth of higher-order interactions (L2A and A2A edges) induced by topological propagation, whereas the number of L2L edges remains unchanged. Prediction performance improves significantly for small $r$ but saturates beyond $r=10$, despite a continued rapid increase in graph complexity.