采用反向构成迭代的视觉SLAM直接法位姿估计

Direct Pose Estimation for Visual SLAM Based on Inverse Compositional Iteration

为了提高视觉SLAM相机位姿估计的计算效率,提出一种直接位姿估计方法。该方法无需逐帧进行特征提取和影像匹配,直接根据像点辐射信息构建非线性优化问题求解相机位姿。推导了基于灰度变换系数的直接法位姿估计平差模型及相应的反向构成迭代解法,利用常速运动模型构建合理初值假设,设计相应的关键帧判断与创建准则,并基于ORB-SLAM系统框架进行实现,使其在保留间接法优点的同时,运算速度得以显著提升。实验结果表明:该方法估计每帧位姿仅需5.7ms,可以在保证同等精度的情况下,达到ORB-SLAM系统2倍以上的整体计算效率,基本满足视频影像实时稳健三维重建的要求。

In order to improve the computational efficiency of pose estimation for visual SLAM, a direct pose estimation algorithm is proposed. The camera pose is directly solved by the construction of nonlinear optimization problem on photometric information without the step of feature extraction or image matching for every frame. The formulas of the direct pose estimation adjustment model and the corresponding inverse compositional iteration solution are deduced based on the grayscale transform parameters in detail. By utilizing the constant-velocity model, the reasonable initial value hypothesis is obtained followed by the design of new key-frame judgment and creation criteria. Implemented on the framework of ORB-SLAM system, the computation speed of the algorithm is significantly increased with retaining the advantages of indirect methods. The experimental results on open-source datasets show that the proposed algorithm can estimate the pose of each frame in only 5.7 ms, and can achieve more than twice the overall computational efficiency of ORB-SLAM system with the same accuracy, which basically meets the requirements of video images’ real-time and robust 3D reconstruction.