Active object detection (AOD) enables a system to actively adjust camera parameters or plan the next viewpoint to improve detection accuracy when the current visual input is insufficient. However, most existing AOD methods assume that the target object is visible from the initial viewpoint, which is often unrealistic and reduces task efficiency. To address this limitation, we propose a novel AOD framework that leverages partial prior information to enhance detection performance and task efficiency. Specifically, we construct an extensible prior information library that describes large and easily identifiable adjacent objects (Adj-objects) that are spatially related to the target. This allows the system to initiate AOD based on the presence of an Adj-object, even when the target is initially out of view. Our approach incorporates a duelling deep Q-learning network (Duelling-DQN) with a newly designed reward function to effectively utilise prior information. Additionally, we introduce a viewpoint storage scheme to support fast retrieval and transition between viewpoints. We evaluate the proposed method on the Active Vision Dataset (AVD) and compare it with several state-of-the-art (SOTA) approaches. The experimental results show that our method achieves a superior average success rate of 81.3%, demonstrating its effectiveness in overcoming the initial state limitations of traditional AOD tasks.