基於單目視覺之社會群體交互與視角自適應的行人軌跡預測模型

Abstract

行人作為智慧型運輸系統中一重要的參與者，其行為影響交通安全系統與自動駕駛控制的反應決策。本研究提供了一種端對端的社會交互模型來達成即時的行人軌跡預測，適用於智慧型運輸系統（Intelligent Transportation Systems , ITS）中影像設備的建置環境。透過BoxMOT框架串接的多物件連續追蹤模組保證了即時運算與模型輕量化的能力，本研究基於此追蹤結果接續建立行人的社會交互特徵並透過深度網路預測未來可能路徑。本研究特別針對現實場域與預設視角不同而產生的多視角挑戰，提出一套視角自適應的視覺策略，能夠重建行人的相對位置關係，旨在簡化實際部署時座標重校正程序，使此套方法在不同相機傾角下仍可維持穩定的資料分析能力。此外，針對行人社會交互特徵的學習策略，本研究進一步改良 Social-STGCNN 架構，導入群體標註資料，強化模型在進行單一行人預測時對群體行為的理解能力。Pedestrians, as key participants in Intelligent Transportation Systems (ITS), significantly influence the decision-making processes of traffic safety systems and autonomous driving control. This research proposes an end-to-end social interaction model for real-time pedestrian trajectory prediction, tailored for deployment in ITS environments equipped with visual sensors. Leveraging the BoxMOT framework, the integrated multi-object tracking module ensures real-time performance and model efficiency. Building upon these tracking results, the model extracts social interaction features among pedestrians and employs a deep neural network to predict future trajectories.To address the challenge of multi-view discrepancies arising from varying real-world camera angles, we introduce a view-adaptive visual strategy that reconstructs relative pedestrian positions. This approach simplifies coordinate recalibration during deployment and maintains robust analytical performance across diverse camera perspectives. Furthermore, this research enhances the Social-STGCNN architecture by incorporating group annotation data, thereby improving the model’s capacity to understand group behavior during individual trajectory prediction.