动态环境下基于深度学习的实时视觉SLAM

Real-time visual SLAM based on deep learning in dynamic environment

针对动态环境下视觉同步定位与建图(SLAM)系统运行效率低的问题,提出了一种基于深度学习的实时语义视觉SLAM算法。所提算法的语义线程不会阻塞跟踪线程的运行,只对最新的待分割图像进行语义分割,分割后对分割结果进行检测与修复,并更新语义信息。图像输入后,首先,更新待分割的图像;然后,利用最新的语义信息和L-K光流法剔除动态关键点,并用剩余的特征点对位姿进行估计;最后,利用语义信息和跟踪信息构建出语义点云地图。在TUM数据集上对所提算法进行实验,实验测试结果表明:高动态环境下,所提算法相较于DS-SLAM在绝对轨迹误差上减小了12.54%~91.78%,跟踪一帧的平均耗时为25.91 ms,验证了所提算法在高动态环境中有较高的定位精度以及实时性,相较于无检测修复的算法在高动态环境下拥有更好的建图效果。

Aiming at the low efficiency of visual Simultaneous Localization And Mapping(SLAM)system in dynamic environment,a real-time semantic visual SLAM algorithm based on deep learning was proposed,in which the operation of the tracking thread was not blocked by the semantic thread of the algorithm,and semantic segmentation was only performed on the latest image to be segmented.After segmentation,the segmentation result was checked and repaired,and the semantic information was updated.After the image was input,firstly,the image to be segmented was updated.Then,the dynamic key points were eliminated by using the latest semantic information and L-K(Lucas-Kanade)optical flow method,and the pose was estimated by using the remaining feature points.Finally,the semantic point cloud map was established by using semantic information and tracking information.The experiment was carried out on TUM dataset.The experimental results show that the proposed algorithm decreases the absolute trajectory error by 12.54%-91.78%compared with DSSLAM(Dynamic and Semantic SLAM)in high dynamic environment,and the average time of tracking a frame is 25.91 ms.It is verified that the algorithm has high positioning accuracy and real-time performance in high dynamic environment,and the proposed algorithm has better mapping effect in high dynamic environment than the algorithm without check and repair.