Cross-View Online Action Detection (CV-OAD) aims to learn action feature representations from videos of known viewpoints and generalize to unseen viewpoints for real-time action detection. It holds significant application value in intelligent surveillance, multi-camera systems, and related fields. However, existing methods face challenges in handling viewpoint discrepancies, feature domain adaptation, and temporal modeling. This paper proposes a novel Contrastive Learning with Future frame Guided method for cross-view Online Action Detection (CLFG-OAD). The method employs a GRU architecture to process inputs from different viewpoints and innovatively introduces a future frame-guided contrastive learning strategy during the training phase. Specifically, we design a Future-Guided Contrastive Learning loss function, which enhances feature representation quality through triplet contrastive learning, effectively suppressing background frame interference. Experiments on the IKEA and DAHLIA datasets demonstrate that our method significantly outperforms state-of-the-art approaches. Furthermore, ablation studies confirm the contribution of each component to the model’s cross-view performance, providing a new technical direction for cross-view online action detection. The related code is available at https://github.com/LuYang-2023/ISAIR_2025.git .