Research map / Graph learning / Knowledge and temporal graphs
Higher-order Relational Reasoning for Pedestrian Trajectory Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HighGraph, a plug-and-play graph module with a collision-aware kernel and higher-order adjacency powers that aggregates social influences from different social depths to improve pedestrian trajectory predictors.
Problem
Pedestrian trajectory prediction typically models only pair-wise or group-wise social relations, ignoring indirect higher-order influences that propagate through socially distant agents.
Approach
A collision-aware kernel weights agent pairs by their potential collision point, and higher-order graph convolution (powers of the adjacency matrix) with learnable combination of features across degrees captures chained social influences; the module plugs into existing predictors.
Key results
- Plugging HighGraph into Social-STGCNN reduces ETH/UCY average minADE20/minFDE20 from 0.44/0.75 to 0.39/0.59 (11%/21% relative).
- On SDD, HighGraph improves Social-GAN minADE20/minFDE20 from 27.23/41.44 to 18.60/37.00 and Social-STGCNN from 20.60/33.10 to 15.38/26.41.
- Gains on CVAE-based models are smaller, e.g., SocialVAE ETH/UCY average goes from 0.21/0.33 to 0.20/0.30.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add HighGraph | SocialVAE · ETH/UCY AVG | minADE20: 0.21 -> 0.20 | Helped |
| add HighGraph | SocialVAE · ETH/UCY AVG | minFDE20: 0.33 -> 0.30 | Helped |
| add HighGraph | EigenTrajectory · SDD | minADE20: 8.05 -> 7.81 | Helped |
| add HighGraph | EigenTrajectory · SDD | minFDE20: 13.25 -> 11.09 | Helped |
| use a collision-aware kernel with distance scaling rather than no weights (ablation) | Social-STGCNN · ETH/UCY AVG | minADE20: 0.49 -> 0.39 | Helped |
| use a collision-aware kernel with distance scaling rather than no weights (ablation) | Social-STGCNN · ETH/UCY AVG | minFDE20: 0.79 -> 0.61 | Helped |
| use a collision-aware kernel with distance scaling rather than a velocity-based kernel (ablation) | Social-STGCNN · ETH/UCY AVG | minADE20: 0.44 -> 0.39 | Helped |
| use a collision-aware kernel with distance scaling rather than a velocity-based kernel (ablation) | Social-STGCNN · ETH/UCY AVG | minFDE20: 0.75 -> 0.61 | Helped |
| add GCN relational reasoning (ablation) | LSTM encoder-decoder · ETH/UCY AVG | minADE20: 1.32 -> 1.31 | Helped |
| use MixHop rather than GCN for relational reasoning (ablation) | LSTM encoder-decoder · ETH/UCY AVG | minADE20: 1.31 -> 1.30 | Helped |
| use HighGraph rather than GCN for relational reasoning (ablation) | LSTM encoder-decoder · ETH/UCY AVG | minADE20: 1.31 -> 1.19 | Helped |
| use HighGraph rather than MixHop for relational reasoning (ablation) | LSTM encoder-decoder · ETH/UCY AVG | minFDE20: 1.86 -> 1.75 | Helped |
Surprising findings
- Improvements are smallest on the UNIV subset, attributed to a train/test mismatch in pedestrian count per sequence.
- Gains on CVAE-based models are minor relative to non-CVAE baselines.
Limitations
- Hyperparameters (layers and degree M) are searched only within {1..5}, and larger values deteriorate performance.
- Benefits are smallest on UNIV, where test sequences have far more pedestrians than training sequences.
- Qualitative multi-agent higher-order examples were manually selected.
Experimental setup
Datasets: ETH/UCY Stanford Drone Dataset
Benchmarks: ETH/UCY Stanford Drone Dataset
Metrics: minADE20 minFDE20
Baselines: Social-GAN SoPhie Social-STGCNN BiTraP SocialVAE EigenTrajectory GCN MixHop
Builds on: Social-STGCNN GCN MixHop SocialVAE EigenTrajectory
Closest papers by meaning
- A Set of Control Points Conditioned Pedestrian Trajectory Prediction (AAAI 2023 · similarity 0.90)
- GroupNet: Multiscale Hypergraph Neural Networks for Trajectory Prediction With Relational Reasoning (CVPR 2022 · similarity 0.89)
- SocialMOIF: Multi-Order Intention Fusion for Pedestrian Trajectory Prediction (CVPR 2025 · similarity 0.88)
- SocialCircle: Learning the Angle-based Social Interaction Representation for Pedestrian Trajectory Prediction (CVPR 2024 · similarity 0.87)
- SocialCVAE: Predicting Pedestrian Trajectory via Interaction Conditioned Latents (AAAI 2024 · similarity 0.86)
- MS-TIP: Imputation Aware Pedestrian Trajectory Prediction (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.