An active object detection model with multi-step prediction based on deep q-learning network and innovative training algorithm

Active Object Detection (AOD) gathers additional information by deliberately adjusting the agent’s viewpoint, ensuring precise detection results in complex environments. Viewpoint planning(VP) is one of the focal points of attention in AOD. Until now, the predominant approach in implementing AOD algorithms has involved the use of deep q-learning networks(DQNs), with a single discrete action as the output. Nevertheless, these methods exhibit shortcomings in both implementation efficiency and success rate. To address these challenges, an AOD algorithm is proposed in this paper, allowing for multistep prediction and employing a novel training strategy. In more detail, the AOD network using a shared decision-making approach is first constructed, simultaneously outputting the action category and range. Moreover, a novel training method based on the Prioritized Experience Replay(PER) is introduced in this article, enhancing the operational success rate of the AOD algorithm. Finally, the reward function is optimized for the designed framework, thereby promoting the convergence of network training. Several comparable methods are tested on a public dataset(Active Vision Dataset), and the results clearly illustrate the superiority of the approach presented in this article.