In this paper, we propose a key object detection network (KODnet) with RL. This study could reduce the dependence on text data for attention object analysis to some extent. Simulating the attention transfer mechanism, the model can keep an eye on old key objects while switching scenes rapidly and at the time detect multiple key objects quickly in the current timestamp. To the best of our knowledge, there are few studies on detecting key objects by simulating the attention transfer mechanism with RL. From the state-of-the-art comparison experiment results, it indicates that our model is about 6.9% higher than the best performance of the advanced methods. Furthermore, our method can maintain a good AP rate in the case of high IoU thresholds. In this work, there is still much space for improving the model computation speed and optimizing the model structure. In the future, we plan to explore more effective ways to integrate deep learning models with reinforcement learning strategies.