Research map / Video / Video generation
Continuous Locomotive Crowd Behavior Generation
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.
Introduces CrowdES, a diffusion-based crowd emitter plus state-switching (Markov-chain) crowd simulator that generates continuous, controllable crowd trajectories from a single scene image.
Problem
Existing crowd generation methods synthesize momentary scenes or fixed-window trajectories and cannot reproduce the continuous, evolving nature of real-world crowds; no benchmark exists for long-horizon crowd generation realism.
Approach
A particle-system-inspired diffusion emitter places agents on a timeline conditioned on predicted spatial layouts (segmentation, appearance, density maps, population probability), and a simulator uses a navigation mesh plus a Markov-chain switching dynamical system over behavior clusters to produce long-term trajectories; the two alternate in windows to generate lifelong crowds.
Key results
- CrowdES outperforms SE-ORCA and VAE on most scene-level realism and agent-level metrics across ETH, HOTEL, UNIV, ZARA1, ZARA2, SDD, GCS and EDIN datasets.
- On SDD, the diffusion emitter reduces density EMD to 0.038 versus 0.052 for a VAE variant (ablation).
- CrowdES generates a one-hour scenario in about 48 seconds; it has slightly more collisions than SE-ORCA (e.g., 0.411 vs 0.045-0.011 range on several sets).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CrowdES instead of SE-ORCA | Crowd behavior generation · ETH | DTW: 2.122 -> 1.649 | Helped |
| use CrowdES instead of SE-ORCA | Crowd behavior generation · HOTEL | DTW: 0.648 -> 0.643 | No clear effect |
| use CrowdES instead of SE-ORCA | Crowd behavior generation · UNIV | DTW: 1.279 -> 1.121 | Helped |
| use CrowdES instead of SE-ORCA | Crowd behavior generation · ZARA1 | DTW: 1.256 -> 0.675 | Helped |
| use CrowdES instead of SE-ORCA | Crowd behavior generation · ZARA2 | DTW: 1.244 -> 0.579 | Helped |
| use CrowdES instead of SE-ORCA | Crowd behavior generation · SDD | DTW: 6.645 -> 6.352 | Helped |
| use CrowdES instead of SE-ORCA | Crowd behavior generation · GCS | DTW: 3.289 -> 3.864 | Hurt |
| use CrowdES instead of SE-ORCA | Crowd behavior generation · EDIN | DTW: 1.819 -> 1.361 | Helped |
| add diffusion-based generation (ablation) | Crowd emitter on SDD · SDD | Dens.: 0.052 -> 0.038 | Helped |
| add population sampling (ablation) | Crowd emitter on SDD · SDD | Freq.: 0.042 -> 0.033 | Helped |
| add spatial layout condition (ablation) | Crowd emitter on SDD · SDD | Dens.: 0.039 -> 0.038 | Helped |
| add collectivity transformer (ablation) | Crowd emitter on SDD · SDD | Freq.: 0.037 -> 0.033 | Helped |
Surprising findings
- CrowdES has more collisions than the ORCA-based SE-ORCA baseline on several datasets, attributed to its focus on behavior switching.
Limitations
- Slightly higher collision rate than SE-ORCA on several datasets.
- Visualizations use CARLA-synthesized virtual crowds rather than real renderings.
Experimental setup
Datasets: ETH UCY SDD GCS EDIN
Benchmarks: ETH UCY SDD GCS EDIN
Metrics: Density EMD Frequency EMD Coverage EMD Population EMD Kinematics EMD DTW Diversity Collision rate
Baselines: SE-ORCA VAE
Builds on: SegFormer Grounded-SAM ORCA Recast navigation mesh DDIM
Closest papers by meaning
- Social Physics Informed Diffusion Model for Crowd Simulation (AAAI 2024 · similarity 0.85)
- PACER+: On-Demand Pedestrian Animation Controller in Driving Scenarios (CVPR 2024 · similarity 0.84)
- Learning Extremely High Density Crowds as Active Matters (CVPR 2025 · similarity 0.83)
- TIMotion: Temporal and Interactive Framework for Efficient Human-Human Motion Generation (CVPR 2025 · similarity 0.83)
- CrowdDiff: Multi-hypothesis Crowd Density Estimation using Diffusion Models (CVPR 2024 · similarity 0.83)
- Learning to Generate Diverse Pedestrian Movements from Web Videos with Noisy Labels (ICLR 2025 · similarity 0.83)
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.