<p>Open intent detection is a critical task for dialog systems, designed to accurately classify known intents into their respective categories and also detect unknown intents. In this paper, we propose a framework of multi-path feature fusion and constrained boundary optimization for open intent detection (FFBO-OID). First, we finely tune a BERT model using gated weights to integrate hierarchical and global features and operate a joint loss training scheme to enhance the adaptability of the model for class-unbalanced datasets. Second, we introduce constraint factors for both positive and negative samples and learn a more appropriate decision boundary for each known intention to enhance the generalization capability of the model. Finally, the proposed method shows substantial performance improvements and stability across varying proportions of known categories on three benchmark datasets, outperforming other state-of-the-art methods.</p>