Active Camera View Planning Based on Deep Reinforcement Learning in Indoor Exploration

In the last decades, significant progress has been made in the field of active vSLAM (visual simultaneous localization and mapping). However, the majority of active vSLAM methods suffer from the problem of coupling the camera's viewpoint to the robot's motion, which limits the initiative during the exploration of unknown environments. To address this issue, this study proposes a camera view planning method based on deep reinforcement learning (DRL), which makes the camera rotate independently. First, feature encoding is performed on the reconstructed navigation map using a multi-layer convolutional neural network (CNN). Second, the DRL-based camera view planning method is trained through the proximal policy optimization (PPO) algorithm. Last, a significant number of tests is performed on a public dataset of indoor environments. The simulation results show that the proposed camera view planning method improves the exploration coverage by an average of 4% (33.86 m² for traditional vs. 35.21 m² for the suggested method). Meanwhile, the kind of semantic information newly observed in each step increases by 21% (2.9 for traditional vs. 3.5 for the proposed method) on average, indicating the improvement in the efficiency of information acquisition.