Abstract: To address the issues in existing zero-shot object navigation methods, where large language models suffer from high computational complexity and fail to meet real-time requirements, and object-relation-based modeling methods are susceptible to environmental changes and exhibit poor generalization ability, a zero-shot object navigation method based on open-vocabulary object detection is proposed. To meet the dual demands of real-time performance and target recognition accuracy in open-world navigation scenarios, an object detection optimization strategy oriented to navigation tasks is constructed, where a spatial constraint and confidence fusion mechanism is introduced into the open-vocabulary object detection model to achieve reliable object detection under low latency. To enhance the agent perception for target semantics and environmental space, a target-oriented semantic extraction network with a feedback mechanism is designed. By using a cross-attention mechanism to extract the association information between target semantics and environmental features, the agent is enabled to actively perceive the current state and make timely adjustments, thereby avoiding ineffective exploration. Experimental results show that, compared with TDANet, the proposed method improves the accuracy of unknown object search by 19.5% and 22.0% under the 18/4 and 14/8 category splits, respectively, and improves the accuracy of known object search by 6.6% and 6.3%, respectively. In addition, experimental validation in real-world environments further demonstrates the adaptability of the proposed method to different environments and its generalization ability when facing unknown targets.