TSCA-Net: New AI achieves record pedestrian trajectory prediction accuracy

Pedestrian trajectory prediction in crowded spaces just got a serious upgrade. Researchers from multiple institutions introduced TSCA-Net, a framework that tackles the inherent multimodality of human motion and varying scene complexity through three complementary modules. The Temporal-Spatial Clique Attention (TSCA) module introduces learnable temporal gating into clique-based goal-history interactions, allowing the model to dynamically weigh historical observations relative to each candidate goal. The Cross-Pedestrian Clique Potential (CPCP) module models asymmetric pairwise agent relationships using a dynamic clique potential framework with a time-varying social graph, capturing how each person's movement influences others differently over time.

Rounding out the architecture is the Adaptive KAN Grid Refinement (AKGR) mechanism, which adjusts the B-spline grid resolution of a Kolmogorov-Arnold Network-augmented LSTM decoder based on per-agent goal distribution entropy. This balances model expressiveness against overfitting across varying motion complexities. Tested on ETH/UCY and Stanford Drone Dataset (SDD), TSCA-Net achieves state-of-the-art results with average ADE/FDE of 0.13/0.20 meters on ETH/UCY and 6.95/10.43 pixels on SDD—significantly beating prior methods. Ablation studies confirm each module contributes uniquely to overall performance. The paper has been submitted to IEEE ICDM 2026 Applied Track.