Zhao, Xiucong (Xi'an Jiaotong University), Liu, Shuai (Xi'an Jiaotong University), Qin, Yechen (Beijing Institute of Technology), Lin, Chenhao (Xi'an Jiaotong University), Shen, Chao (Xi’an Jiaotong University)

Temporal Spatial Multi-Scale Compressor Trajectory Prediction Network

Scheduled for presentation during the Regular Session "S31c-AI-Driven Motion Prediction and Safe Control for Autonomous Systems" (FR-LA-T31), Friday, November 21, 2025, 16:40−17:00, Southport 1

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Precise prediction of surrounding-agent trajectories is a cornerstone for ensuring safety and efficiency in autonomous driving. Yet, as traffic density and scene complexity grow, Transformer-based models—with their quadratic attention cost—struggle to operate in real time without sacrificing accuracy. To bridge this gap, we introduce Temporal-Spatial Multiscale Compression (TSMC), a novel framework that hierarchically compresses input features across both temporal and spatial dimensions, effectively ``zooming out'' on less critical interactions while retaining full resolution for salient ones. By reducing the number of tokens processed at each stage, TSMC lowers the computational load when handling large numbers of traffic participants and complex contexts. We integrate TSMC into several state-of-the-art trajectory predictors and evaluate on two popular benchmarks (Argoverse I and Argoverse II). Results show consistent gains in prediction accuracy and a speed-up in inference, with negligible model-size increase. TSMC thus offers a flexible, plug-and-play solution for long-horizon, high-density trajectory forecasting in real-world autonomous systems.