With the advancement of robotics technology, service robots are increasingly being applied in object search tasks to improve work efficiency. However, conventional object search methods still face limitations in real-world applications, particularly in handling unknown and dynamic environments. These environments impose higher demands on search accuracy and efficiency. In this paper, we recommend a scene graph-enhanced embodied decision-making method for autonomous object search, aiming to enhance the robot’s object search capabilities in unknown and dynamic environments. The method primarily incorporates two modules: firstly, a global-scale scene graph is incrementally constructed to semantically sense the environment. Secondly, we propose an embodied decision-making approach guided by large language models (LLMs) to direct the robot in searching for target objects, which enhances the accuracy and efficiency of object search. Our approach is validated via simulations, and the results show that the method performs well in fast and precise object search.