Training-free pedestrian trajectory prediction via segmentation-guided path planning

Abstract

Pedestrian trajectory prediction is a critical research topic for industrial applications and has been significantly advanced by deep learning. Despite decades of progress, current approaches still face two major challenges. First, the scarcity of data limits the generalization capability of deep learning models. Second, the absence of interpretability hinders real-world applications. To address these challenges, recent research has leveraged the knowledge of large language models (LLMs) to alleviate data sparsity and introduced novel knowledge-based methods to enhance interpretability. Nevertheless, transferring LLMs is excessively cumbersome, and the black-box nature of deep learning continues to obstruct interpretability. In our work, we propose a training-free transfer paradigm named Segmentation-Guided Path Planning (SGPP). Rather than directly transferring pretrained LLMs, SGPP introduces a more tailored and efficient transfer strategy by employing a promptable segmentation model. The segmentation model explicitly extracts walkable regions from the scenario, which serve as constraints in the planning space, thereby reformulating trajectory prediction as a more tractable white-box path-planning problem. Within this framework, our method offers a more effective solution to the two prevailing challenges. Compared with the latest training-free methods, our approach achieves superior performance and demonstrates strong generalization across diverse real-world scenarios, highlighting its suitability for industrial applications.