在室外等大规模环境下,由于存在着光照、 遮挡、 大量相似物和相机的旋转、 抖动等问题,传统的基于特征提取的方法会在实时性、 数据关联、 闭环检测等方面存在限制.文中在视觉测程法的基础上使用图像直接对齐,并结合基于滤波的方法进行半稠密深度图估计,进而把选取出来的关键帧图像融合到全局地图中,最后在后端采用改进的g2o[1]框架不断地对位姿图进行优化.实验结果表明,该方法取得了比较好的实时性和鲁棒性.%In large-scale outdoor environment, there are light, block, a number of similar material and the rotation of the camera, jitter and so on, the traditional method based on feature extraction is limited about real-time, data association and loop closure detection. This paper uses direct image alignment based on direct visual odometry accoupled with filtering-based estimation of semi-dense depth maps, then fusion the selected keyframes into global map, at last,it uses advanced g2o framework to optimize the pose-graph in the back-end. The result show that this method achieved a better real-time and robustness.
展开▼
