Research map / Multi-agent systems and games / Human-AI and goal-conditioned
SocialCircle: Learning the Angle-based Social Interaction Representation 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 SocialCircle, an angle-based, echolocation-inspired social interaction representation (velocity, distance, direction per angular partition) fused into trajectory prediction backbones to improve pedestrian forecasting.
Problem
Modeling social interactions among agents for pedestrian trajectory prediction is hard: model-based rules are hard to generalize and model-free networks lack explainability.
Approach
Encode neighbors' relative positions into angular partitions around the target agent, with per-partition average velocity, distance and direction, and treat this cyclic spatial representation as a virtual temporal sequence fused with the observed trajectory embedding in existing backbones.
Key results
- E-V2-Net-SC reaches 0.17/0.27 ADE/FDE on ETH-UCY, matching the best reported baseline value and improving over Agentformer by 5.6% ADE and 6.9% FDE.
- V2-Net-SC improves over Y-net on SDD by 14.52% ADE and 10.04% FDE (6.71/10.66 vs 7.85/11.85).
- Adding SocialCircle to the simple Transformer improves ADE by 5.89% and FDE by 4.00% on SDD, and lowers training-loss NaN failures.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SocialCircle to E-V2-Net | E-V2-Net · ETH-UCY | FDE: 0.28 -> 0.27 | Helped |
| Add SocialCircle to V2-Net | V2-Net · SDD | ADE: 7.12 -> 6.71 | Hurt |
| Add SocialCircle to E-V2-Net | E-V2-Net · NBA Metrics@2.0s | ADE: 0.68 -> 0.67 | Helped |
| Add SocialCircle to E-V2-Net | E-V2-Net · NBA Metrics@4.0s | ADE: 1.26 -> 1.18 | Helped |
| Add velocity component (ablation) | V2-Net-SC · SDD | ADE | Helped |
| Add distance component (ablation) | V2-Net-SC · SDD | ADE | Helped |
| Add direction component (ablation) | MSN-SC · SDD | FDE: up to 3.30% FDE gain | Helped |
| Add velocity component (ablation) | MSN-SC · SDD | ADE: less than 1.07% ADE differences | No clear effect |
| Add SocialCircle to Transformer | Transformer · SDD | ADE: 5.89% ADE | Helped |
Surprising findings
- Five of six plain Transformer training runs diverged to NaN loss before 450 epochs, while only two SocialCircle runs did, suggesting SocialCircle improves training stability.
Limitations
- Does not encode the direction from which neighbors come.
- Does not explicitly model higher-order interactions among neighbors.
- Gains over some strong baselines are mixed; MSN-SC is worse than some baselines on ETH-UCY and NBA ADE is worse than GroupNet+CVAE.
Experimental setup
Datasets: ETH-UCY Stanford Drone Dataset NBA SportVU
Benchmarks: ETH-UCY SDD NBA SportVU
Metrics: minADE20 minFDE20
Baselines: SHENet MID EqMotion MSN LED Trajectron++ Agentformer V2-Net Y-net E-V2-Net FlowChain IMP
Builds on: Social Pooling Transformer MSN V2-Net E-V2-Net
Closest papers by meaning
- SocialCVAE: Predicting Pedestrian Trajectory via Interaction Conditioned Latents (AAAI 2024 · similarity 0.89)
- SocialMOIF: Multi-Order Intention Fusion for Pedestrian Trajectory Prediction (CVPR 2025 · similarity 0.88)
- Social-Transmotion: Promptable Human Trajectory Prediction (ICLR 2024 · similarity 0.88)
- Higher-order Relational Reasoning for Pedestrian Trajectory Prediction (CVPR 2024 · similarity 0.87)
- Following the Human Thread in Social Navigation (ICLR 2025 · similarity 0.85)
- A Unified Environmental Network for Pedestrian Trajectory Prediction (AAAI 2024 · similarity 0.85)
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.