Research map / Multi-agent systems and games / Human-AI and goal-conditioned
SocialMOIF: Multi-Order Intention Fusion 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 SocialMOIF, a pedestrian/agent trajectory predictor that fuses first-order and higher-order neighbor intentions, guides latent trajectory distributions with a approximator, and uses a KAN-based global optimizer for parallel prediction with a distance-direction loss.
Problem
Short-term agent trajectory prediction is hard due to uncertain agent intentions and higher-order influences among neighboring agents, and sequential decoding accumulates errors and is slow.
Approach
Combines a first-order intention layer (target-neighbor interactions) with a higher-order layer (neighbor-neighbor interactions) into a multi-order fused intention, uses a latent-variable trajectory distribution approximator supervised by future trajectories, introduces Kolmogorov-Arnold networks (KANs) for global parallel trajectory optimization, and adds a distance-direction loss.
Key results
- Reduces ADE/FDE on ETH/UCY average to 0.13/0.18 m (best-of-20), beating E-V2-Net-SC at 0.15/0.20.
- On NBA Rebound, reports ADE/FDE of 0.34/0.66 m versus prior SOTA 0.54/0.79; on SDD, 0.17/0.24 m.
- Ablation on SDD shows the full model drops ADE from 0.46 to 0.17 relative to a baseline with only position and distance inputs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add neighbor velocities (V) (ablation) | SocialMOIF PDI baseline · SDD | ADE: 0.46 -> 0.45 | Helped |
| Add neighbor velocities (V) (ablation) | SocialMOIF PDI baseline · NuScenes | ADE: 1.26 -> 1.24 | Helped |
| Add interaction angle (Θ) (ablation) | SocialMOIF PDI baseline · SDD | ADE: 0.45 -> 0.41 | Helped |
| Add interaction angle (Θ) (ablation) | SocialMOIF PDI baseline · NuScenes | ADE: 1.24 -> 1.18 | Helped |
| Add final historical neighbor distance (E) and global historical social intent (I) (ablation) | SocialMOIF · SDD | ADE: 0.41 -> 0.35 | Helped |
| Add final historical neighbor distance (E) and global historical social intent (I) (ablation) | SocialMOIF · NuScenes | ADE: 1.18 -> 1.12 | Helped |
| Add future trajectory approximator (B) (ablation) | SocialMOIF · SDD | ADE: 0.35 -> 0.30 | Helped |
| Add future trajectory approximator (B) (ablation) | SocialMOIF · NuScenes | ADE: 1.12 -> 1.04 | Helped |
| Add global parallel optimizer (K) (ablation) | SocialMOIF · SDD | ADE: 0.30 -> 0.20 | Helped |
| Add global parallel optimizer (K) (ablation) | SocialMOIF · NuScenes | ADE: 1.04 -> 0.93 | Helped |
| Add directional angle loss (A) (ablation) | SocialMOIF · SDD | ADE: 0.20 -> 0.17 | Helped |
| Add directional angle loss (A) (ablation) | SocialMOIF · NuScenes | ADE: 0.93 -> 0.92 | Helped |
Limitations
- Gains on ETH/UCY are smaller than on NBA and SDD, where interactions are more complex.
- Only short-term prediction (8 future frames from 8 history frames) is evaluated.
- Interpretability claims rest on qualitative heatmaps rather than quantitative measures.
Experimental setup
Datasets: ETH UCY SDD NBA 2015-2016 NuScenes
Benchmarks: ETH/UCY NBA SDD NuScenes
Metrics: ADE FDE NLL
Baselines: FLEAM ST-MR MTN MANTRA Agent-former BiTrap MemoNet EqMotion V2-Net-SC E-V2-Net-SC SGNet Y-Net
Builds on: V2-Net EqMotion Kolmogorov-Arnold networks CVAE
Closest papers by meaning
- Interpretable Social Anchors for Human Trajectory Forecasting in Crowds (CVPR 2021 · similarity 0.89)
- SocialCVAE: Predicting Pedestrian Trajectory via Interaction Conditioned Latents (AAAI 2024 · similarity 0.89)
- Higher-order Relational Reasoning for Pedestrian Trajectory Prediction (CVPR 2024 · similarity 0.88)
- Multimodal Interaction-Aware Trajectory Prediction in Crowded Space (AAAI 2020 · similarity 0.88)
- SocialCircle: Learning the Angle-based Social Interaction Representation for Pedestrian Trajectory Prediction (CVPR 2024 · similarity 0.88)
- Social-Transmotion: Promptable Human Trajectory Prediction (ICLR 2024 · similarity 0.88)
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.