基于改进关键帧选择的ORB-SLAM2算法

针对同时定位与建图（SLAM）算法精度不高且跟踪易失败的问题，提出了一种改进关键帧选择的ORB-SLAM2算法。通过ORB-SLAM2算法计算帧间相对位姿；在原有算法的基础上，增加旋转与平移量作为判定依据，决定是否创建新关键帧；针对移动机器人所安装的相机与机器人产生相对运动引发误拍摄，导致劣质关键帧生成的问题，设计了劣质关键帧剔除算法；基于RGB-D数据集与自主研发的移动机器人进行了实验验证。实验结果表明:改进的关键帧选择算法能够准确及时地选择关键帧，最优情况下定位误差约为原误差的51.9%，有效消除了相机与机器人之间相对运动产生的影响，直线误差仅为原误差的82.1%。改进算法能够有效提高定位精度，减少跟踪失败。 

ORB-SLAM2 algorithm based on improved key frame selection

To address the difficulties caused by the low accuracy and poor robustness of simultaneous localization and mapping (SLAM), an ORB-SLAM2 algorithm is proposed based on key frame selection. First, the relative pose between frames is calculated based on ORB-SLAM2. Second, to determine whether a new key frame should be created, rotation and translation values are added to the original algorithm, functioning as the judgement basis. Then, an inferior key frame removal algorithm is designed to solve the problem of inferior key frame generation which results from incorrect shooting caused by the relative movement between the robot and the camera installed in the self-developed mobile robot. Finally, experiments are carried out based on the RGB-D dataset and the developed mobile robot, verifying the outstanding performance of the proposed algorithm. The results show that the improved key frame selection algorithm can accurately and timely choose the key frame, and reduce tracking failures. In the most optimal case, the positioning error is about 51.9% of that of the original, while the linear error is about 82.1% of that of the original, which effectively eliminates the influence caused by relative motion between the camera and the robot. This research shows that the improved algorithm could effectively promote positioning accuracy and reduce tracking failures.