融合惯性信息的单目直接法定位与稠密地图构建

Robot Localization and Dense Map Construction Based on Monocular Direct Method and Inertial Information

针对单目直接法依靠图像梯度进行优化容易陷入局部最优、且难以构建低纹理区域地图的问题,构建高精度IMU预积分模型,将惯性信息融合到图像跟踪过程中,为视觉跟踪提供精确的帧间运动约束及良好的初始化梯度方向信息,构建视觉惯性跟踪模型,提高了单目视觉的定位精度并实现半稠密地图构建;通过超像素图像分割提取出二维图像不同的轮廓位置,提出双重投影匹配算法确定出可靠的超像素与对应的3D空间点,通过RANSAC对低梯度图像区域进行平面拟合以及异常点剔除,完成低纹理区域的地图扩建,实现稠密点云地图的构建。实验结果表明,与传统视觉定位模型相比,直接法与惯性信息融合提高了系统的定位精度,在无GPU加速的情况下能构建精确的稠密三维点云地图。

Traditionally,the monocular direct method uses image gradients for optimization,it is easy to fall into a local optimum,and it is difficult to construct a low-texture area map.Aiming at these problems,a high-precision IMU pre-integration model is constructed to fuse inertial information into image tracking,which provides accurate inter-frame motion constraints and good initialization gradient direction information for visual tracking.The constructed visual inertial tracking model improves the positioning accuracy of monocular vision and realizes the construction of semi-dense maps.Different contour positions of a two-dimensional image are extracted by using superpixel image segmentation,then the antipodal projection matching algorithm is used to determine the superpixels and the corresponding 3 D points.The RANSAC method is used to eliminate abnormal points and fit the low gradient image region to complete the map expansion of the low-texture region.The dense point cloud image is constructed eventually.The experimental results show that compared with the traditional visual positioning model,the direct method fusing inertial information improves the positioning accuracy of the system,and can construct accurate dense 3 D point cloud maps without GPU acceleration.