月面巡视器基于深度学习的即时定位与建图

Simultaneous Localization and Mapping of Lunar Rover Based on Deep Learning

针对载人登月对月面大范围行走探测以及月面巡视器导航定位的要求,提出了一种基于深度学习的视觉即时定位与建图(SLAM)方法。该方法设计了一个全监督的卷积神经网络对单目SLAM建模,减少了传统方法中人工设计特征和根据场景设置各种参数阈值的局限性;同时,利用深度学习模型良好的迁移学习能力,从大量地面数据训练并在少量仿月表面数据微调中得到网络的参数,从图像序列中直接估计平移量和旋转量;此外,引入了三维点云构成的稀疏深度图作为监督,采用光度误差构造的损失函数将深度信息和位姿信息结合,得到位姿估计的精度比肩传统SLAM算法,同时增加了算法对环境的适应性和鲁棒性。实验证明该算法在城市道路环境和仿月表面环境均有较优的性能。

After robotic lunar exploration mission,the human lunar exploration will be the future goal of China.To satisfy the navigation needs of the lunar rover in larger fields and the needs of more precise localization,a supervised convolutional neural network was proposed to model monocular Simultaneous Localization and Mapping(SLAM),thus to reduce the limitations of manual features.Based on the great transfer learning capability of the deep learning model,and with large amount of data to train the parameters of the network,the pose estimation could be extracted from the image sequence directly.In addition,the sparse depth map composed of 3D point clouds was introduced into the network as a supervised term.Photometric error was adopted to combine the depth information and pose information as part of cost function,thus a more accurate pose estimation was obtained and the adaptability and robustness of the algorithm to the environment were increased.Besides,the algorithm demonstrated promising performance in both urban scene and moon-like surface scene.