基于深度学习与特征点速度约束的室内动态SLAM方法

An indoor dynamic SLAM method based on deep learning and feature point velocity constraint

针对动态环境下同时定位与建图(SLAM)算法定位信息不准确、建图偏移严重的问题,提出一种深度学习与特征点速度约束相结合的室内动态SLAM方法。引入基于深度学习的潜在动态目标检测线程与基于ORB-LK光流金字塔的特征点速度计算线程。首先,通过YOLO_v5算法对场景中潜在的动态目标进行识别判断,提供先验信息;然后,利用ORB-LK光流金字塔算法计算图像中特征点的相对速度,根据特征点速度约束判断潜在动态目标区域的真实动态性,更加合理地剔除动态特征点,提高视觉SLAM系统在室内动态环境下的定位精度和鲁棒性。实验结果表明,在TUM数据集中动态子序列下与ORB-SLAM2、Detect-SLAM、DS-SLAM相比,绝对轨迹均方根误差分别平均减少94.74%、51.95%、43.61%。

Aiming at the problem of inaccurate positioning information and serious deviation of simultaneous localization and mapping(SLAM)algorithm in dynamic environment,an indoor dynamic SLAM algorithm combining deep learning and feature point velocity constraint is proposed.A potential dynamic target detection thread based on deep learning and a feature point velocity calculation thread based on ORB-LK optical flow pyramid are introduced.First,the YOLO_v5 algorithm is used to identify the potential dynamic targets in the scene and provide prior information.Then,the ORB-LK optical flow pyramid algorithm is designed to calculate the relative feature point velocity,and real dynamic points are determined according to feature point velocity constraint,making the dynamic feature points elimination more reasonably and improving the positioning accuracy and robustness of the visual SLAM system in indoor dynamic environment.The experimental results show that compared with ORB-SLAM2,Detect-SLAM and DS-SLAM,the absolute trajectory root mean square error(RMSE)is reduced by 94.74%,51.95%and 43.61%respectively under the dynamic sequences of TUM dataset.