一种改进的语义视觉里程计优化算法

An Improved Semantic Visual Odometry Optimization Algorithm

为了优化移动机器人的视觉SLAM过程中的里程计估算问题,利用视觉图像的语义信息优化使得机器人在进行地图构建过程中的位姿估计更为准确,本文提出了一种改进的视觉里程计优化算法。该算法基于语义视觉里程计算法,提出了一种依赖语义特征不变性对视觉里程计进行约束优化的算法。通过构建由语义分割网络检测结果的语义重投影误差损失函数,并对此函数进行优化求解,从而对视觉SLAM中的里程计实现更高精度的优化。经过实验,结合文中所提出的里程计优化算法的视觉SLAM算法在数据集和实际环境中的里程计误差均有优化。

In order to optimize the odometry estimation problem in the visual SLAM process of mobile robots, the semantic information optimization of visual images makes the robot’s pose estimation in the process of map construction more accurate. In this paper, an improved visual odometry optimization algorithm is proposed. Based on the semantic visual odometry algorithm, this algorithm proposes a constrained optimization algorithm for visual odometry relying on the invariance of semantic features. By constructing the semantic reprojection error loss function detected by the semantic segmentation network, and optimizing this function, the odometer in visual SLAM can be optimized with higher precision. After experiments, the visual SLAM algorithm combined with the odometer optimization algorithm proposed in this paper has optimized the odometer errors in both the da-taset and the actual environment.