【摘要】 特征点提取是视觉同时定位与建图(Visual Simultaneous Localization and Mapping, VSLAM)的重要步骤之一，近年来出现的基于深度学习的特征点提取方法通常效率较低，无法满足实时性要求，也不能提供描述符所需的几何不变性.为此，文中提出轻量化特征点及可变形描述符提取网络(Lightweight Keypoint and Deformable Descriptor Extraction Network, LKDD-Net),在主干网络中引入轻量化网络模块，提高特征提取效率.LKDD-Net可同时获取特征点位置和可变形描述符.为了验证LKDD-Net的有效性，设计视觉里程计系统.在HPatches、TUM RGB-D公共数据集上的实验表明，LKDD-Net可在GPU上实时运行，特征点提取时间仅为8.3 ms,同时在各种场景中保持高精度和强鲁棒性，而且其构成的视觉里程计系统性能较优.更多还原

【Abstract】 Keypoint extraction is a crucial step in visual simultaneous localization and mapping(VSLAM). Existing deep learning based keypoint extraction methods suffer from low efficiency and fail to meet real-time requirements. Furthermore, they do not provide the geometric invariance required by descriptors. To address this issue, a lightweight keypoint and deformable descriptor extraction network(LKDD-Net) is proposed. A lightweight network module is introduced in the backbone network to improve the efficiency of feature extraction, and then the deformable convolution module is applied to the descriptor decoder to extract deformable descriptors. LKDD-Net is capable of simultaneously obtaining both keypoint locations and deformable descriptors. To study the effectiveness of LKDD-Net, a visual odometry system based on LKDD-Net is designed. Experiments on HPatches public dataset and TUM public dataset show that LKDD-Net can run in real-time on GPUs with keypoint extraction time being as low as 8. 3 ms, while maintaining high accuracy in various scenarios. The performance of the visual odometry system composed of LKDD-Net is superior to traditional vision and VSLAM systems based on deep learning keypoint extraction. The proposed method successfully tracks all six sequences in TUM public dataset, demonstrating stronger robustness.更多还原