In this study, the goal is to efficiently and actively search for a target object in a previously unknown large-scale environment. To this end, we develop a probabilistic environment model that can utilize spatial commonsense knowledge and environment-specific spatial relations. The model evaluates the merit of exploring each possible viewpoint in the environment to find the target object. Then, the path planning method incorporates the estimated value of these viewpoints and the time cost between them to generate an efficient search path that minimizes the total search time. We also describe a search space reduction method that improves the feasibility of the proposed approach in large-scale environments. To validate the approach, we compare the search times of the proposed method to those of human participants, a coverage-based search and a random search in simulation experiments. The results show that the proposed method can generate search paths with similar search times to those of human participants, while clearly outperforming the coverage-based and random search methods. We also demonstrate the applicability of the approach in real-world experiments in which the robot could find the target object without a single failure case in 70 trials.