基于深度学习和边缘检测的动态场景下鲁棒SLAM

机器人在执行同时定位与地图创建(simultaneous localization and mapping,SLAM)的复杂任务时,容易受到移动物体的干扰,导致定位精度下降、地图可读性较差、系统鲁棒性不足,为此提出一种基于深度学习和边缘检测的SLAM算法。首先,利用YOLOv4目标检测算法获取场景中的语义信息,得到初步的语义动静态区域,同时提取ORB特征点并计算光流场,筛选动态特征点,通过语义关联进一步得到动态物体,利用canny算子计算动态物体的轮廓边缘,利用动态物体以外的静态特征点进行相机位姿估计,筛选关键帧,进行点云叠加,利用剔除动态物体的点云信息构建静态环境地图。本文算法在公开数据集上与ORB_SLAM2进行对比,定位精度提升90%以上,地图可读性明显增强,实验结果表明本文算法可以有效降低移动物体对定位与建图的影响,显著提升算法稳健性。

Robust SLAM in Dynamic Scenes Based on Deep Learning and Edge Detection

When the robot handling the task of simultaneous localization and mapping,it will be impacted by the moving objects.This factor will lead to precision decline,poor map readability and insufficient robustness.An algorithm based on deep learning and edge detection was proposed to handle this.At first,detection of YOLOv4 is adopted to extract semantic information in the scene.After getting basic semantic area of static and dynamic,it extracts ORB feature points and calculates optical flow field,which filters out dynamic feature points,at this point,it can get dynamic objects by semantic relations.Then,it uses canny to calculate the contour of dynamic objects and gets camera pose by static feature points outside dynamic objects,filters out the keyframe,and overlays the point cloud that eliminated dynamic objects to build static environment map.This paper compares the algorithm above with the ORB_SLAM2 on the public datasets,and the results of experiment show that this algorithm can improve the pose estimation by more than 90%and enhance map readability significantly,so it can reduce the effect introduced by the moving objects,enhance the stability significantly.