基于空间约束ICP的改进视觉里程计

An Improved Visual Odometry Based on AICP

针对传统的迭代最近点(Iterative Closest Point,ICP)算法在视觉里程计应用中存在的外点干扰与相机偶然大运动导致的位姿估计不准确或失效问题,提出基于空间约束ICP算法的相机位姿估计方法与帧对模型与帧间估计的切换策略,提高位姿估计准确度和鲁棒性。方法利用彩色深度相机的数学模型和空间约束为每帧图像提取的三维点云中的点赋予不同权值,并在ICP算法起始阶段估计当前帧点云与模型点云重叠率,减少外点对位姿估计的影响。为应对相机的偶然大运动,提出帧间位姿计算与帧对模型位姿计算的切换策略。实验在标准测试数据集下,与典型视觉里程计对比有效提高小规模运动下位姿估计的准确度,同时解决了相机偶然大运动下系统失效问题。

To avoid using prior information and motion assumptions of object and to address the error drift, a sliding window filter based object pose estimation algorithm is proposed. The algorithm requires no prior information and motion assumptions and avoids the error drift. To estimate the pose of an arbitrary object, Gauss-Newton method is used within each window to optimize the pose and structure iteratively. To tackle the error drift, Kalman filter and the strong coupling of pose and structure are explored by filtering the structure of object. Structure filtering increases the accuracy of structure estimation and thus the accuracy of estimated pose is increased without error drift. Simulation shows that the proposed algorithm can eliminate the error drift effectively and enhance the accuracy of existing method. Without prior information and assumption, the proposed algorithm can achieve the close estimation results of state-of-art methods.