Multimodal Pedestrian Trajectory Prediction with Multi-Scale Spatio-Temporal Group Modeling and Diffusion

Objective The rapid development of autonomous driving and social robotics has increased the need for accurate pedestrian trajectory prediction to improve safety and interaction efficiency. Existing group-based methods mainly emphasize local spatial interaction and often overlook latent grouping characteristics across time. This study proposes a multi-scale spatiotemporal feature construction method that separates trajectory shape from absolute spatiotemporal coordinates. This enables the model to capture latent group associations across different temporal intervals. A spatiotemporal interaction three-element encoding mechanism is incorporated to extract dynamic relationships between individuals and groups. By integrating the reverse process length mechanism of diffusion models, the system progressively reduces prediction uncertainty. This approach provides an effective solution for multimodal trajectory prediction in complex, crowded scenes and offers theoretical support for improving the accuracy and stability of long-range trajectory forecasting. Methods The algorithm performs deep modeling of pedestrian trajectories through multi-scale spatiotemporal group modeling across three components: group construction, interaction modeling, and trajectory generation. First, to address the limitations of methods that focus on local spatiotemporal patterns but overlook cross-dimensional latent characteristics, a multiscale trajectory grouping model is developed. Its core design extracts trajectory offsets to represent trajectory shapes, separating motion features from absolute positions. This enables the system to identify latent group associations among agents who follow similar motion patterns across different periods. Second, a spatiotemporal interaction three-element encoding method is proposed. By defining neural interaction strength, interaction categories, and category functions, the method captures detailed individual interactions and the global dynamic evolution of collective behavior. Finally, a Diffusion Model is introduced for multimodal prediction. Through the reverse process length mechanism, the model converges gradually, reduces uncertainty, and transforms a diffuse prediction space into plausible future trajectories. Results and Discussions The proposed model was evaluated against 11 state-of-the-art baselines on the NBA dataset (Table 1). The results show clear advantages in minADE20. It achieves substantial gains over GroupNet+CVAE in long-term prediction tasks, improving minADE20 and minFDE20 by 0.19 and 0.36, respectively, at the 4-second horizon. Although it is slightly inferior to MID in long-term trend prediction, possibly because group dynamics shift rapidly and intensely in NBA scenarios, the model maintains strong instantaneous accuracy. This supports the effectiveness of the multi-scale grouping strategy, which uses historical trajectories to capture complex dynamic interactions. On the ETH/UCY datasets (Table 2), MSGD provides consistent improvements across all five sub-scenes. In the dense and highly interactive UNIV scene, the method exceeds all baselines by leveraging the strengths of multi-scale modeling. Although MSGD is marginally behind PPT in long-distance endpoint constraints, it maintains a lead in minADE20. It also outperforms Trajectory++ in velocity smoothness and directional coherence (std dev: 0.701 2) (Table 3), indicating that the generated trajectories maintain natural smoothness aligned with human motion. Ablation studies verify the independent effects of the diffusion model, spatiotemporal feature extraction, and multi-scale grouping modules (Table 4). Grouping sensitivity analysis on the NBA dataset shows that full-court grouping (group size 11) enhances long-term stability, reducing minFDE20 by 0.026~0.03 at 4 seconds (Table 5). Configurations with group sizes of 5 or 2 further support the importance of team formations and “one-on-one” local offensive and defensive dynamics (Table 6). Diffusion-step and training-epoch sensitivity analysis reveals a complementary relationship: moderate diffusion steps (30~40) refine denoising and improve accuracy, whereas excessive steps may cause overfitting (Table 7). Qualitative visualization confirms that MSGD generates multimodal trajectories with high overlap with ground truth (Fig. 2). Conclusions This study presents a trajectory prediction algorithm that improves performance in two primary ways: (1) it captures pedestrian interactions by extracting spatiotemporal features, and (2) it strengthens collective behavior modeling through multi-scale grouping. Experiments show that the method achieves state-of-the-art performance on the NBA and ETH/UCY datasets, and ablation studies confirm the effectiveness of all modules. Two limitations remain. First, explicit environmental information, such as maps or obstacles, is not yet incorporated. Second, the diffusion model requires substantial computational cost during inference. Future research will address these issues.