Skip to content

A Relevance-Aware, Multimodal Framework for Efficient Trajectory Prediction in Dense Urban Environments

Aarij Mehdi Shah, Muhammad Rehan, Abdullah Ayub Khan, Bismah Siddiqui, Fizza Tariq, Arzoo

Status
Submitted, under review
Date
2025
Tags
Autonomous driving · Trajectory prediction · Efficiency · Graph pruning

  • 52.31%

    edges pruned

    from a fully connected N(N−1) graph

  • 0.3863 m

    minADE5

  • 0.4981 m

    minFDE5

  • 300

    driving scenes

    nuScenes v1.0-mini

Abstract

Trajectory predictors for autonomous driving usually reason over a fully connected graph of every agent in the scene, which gets expensive fast in dense urban traffic. This work adds a deterministic, relevance-aware pruning stage in front of the predictor: complementary geometric constraints decide which interactions actually matter, and the rest are removed before prediction. On 300 complex scenes from nuScenes v1.0-mini, the framework removes 52.31% of interaction edges while keeping displacement errors low, and it is designed to plug into existing deep predictors such as Trajectron++, HiVT and EWTA.

01The idea

Most agents in a busy street don’t affect each other. Instead of letting the model learn that from a fully connected graph, the framework decides it up front, deterministically, and hands the predictor a sparse graph of the interactions that matter.

02Method

  • Motion baseline: Kalman-smoothed constant-velocity (CV) equations, approximating realistic, noisy perception from camera, LiDAR and radar.
  • Multi-stage filter: spatial distance thresholds, directional velocity vectors and overlapping future CV projections combine into a relevance mask for every agent.
  • Plug-and-play: the pruned graph is designed to feed existing predictors such as Trajectron++, HiVT and EWTA.
  • Lightweight by design: Python, pandas and NumPy on consumer hardware, with no heavy CUDA dependency.

03Results

On 300 scenes from nuScenes v1.0-mini, pruning removed 52.31% of interaction edges (47.69% of the topology stays active) with minADE5 of 0.3863 m and minFDE5 of 0.4981 m.

04Status

Submitted for peer review; no decision yet. Written with Aarij Mehdi Shah (corresponding author), Abdullah Ayub Khan, Bismah Siddiqui, Fizza Tariq and Arzoo, Department of Computer Science, Bahria University Karachi.