深度学习下的视觉SLAM回环检测及其算法

Visual SLAM Loop Detection and Algorithm Based on Deep Learning

作为机器人技术领域的研究重点之一,SLAM在无人驾驶、增强现实、虚拟现实等方面有重要应用。视觉SLAM利用连续的相机帧获取信息,完成环境环境感知,而长时间运行视觉SLAM系统会不断累积邻近帧间误差,影响后端优化收敛。针对此问题,提出一种基于深度学习的回环检测方法,使用回环检测模块以减少邻近帧间的误差积累,克服以人工标记特征点算法为基础的传统视觉SLAM回环检测的不足,提高了系统在复杂环境下的检测准确率。经实验验证,算法获得良好的准确率与速率,能够满足视觉SLAM系统的要求。

As one of the research focuses in the field of robotics,SLAM has important applications in unmanned driving,augmented reality,virtual reality and so on.Visual SLAM uses continuous camera frames to obtain information and realize environmental perception.However,running visual SLAM system for a long time will accumulate errors between adjacent frames,which will affect the back-end optimiza-tion convergence.To solve the problem,a loop detection method based on deep learning is proposed,which uses the loop detection module to reduce the error accumulation between adjacent frames,overcomes the shortcomings of traditional visual SLAM loop detection based on the algorithm of manually marking feature points,and improves the detection accuracy of the system in complex environments.The experi-mental results show that the algorithm has good accuracy and speed,and can meet the requirements of visual SLAM system.