利用解耦光流运动场模型的双目视觉里程计位姿优化方法

A stereo visual odometry pose optimization method via flow-decoupled motion field model

针对地面移动测量系统(MMS)和无人驾驶车(AV)平台双目立体相机采集的图像序列进行实时载体位姿估计优化问题,提出利用光流运动场模型的载体位姿与图像光流矢量间关系,将光流矢量解耦为3个平移分量、3个旋转分量和一个深度分量,推导分析了解耦后单分量、组合分量误差对位姿估计的影响,利用仿真和真实数据试验,验证了不同模型下单分量、组合分量误差分离模型的有效性,并结合组合分量误差分离模型,提出了双目视觉里程计位姿估计的解耦光流运动场位姿优化算法。试验结果表明:该算法可在与初始估计几乎同等计算效率条件下,将载体横向平移平均误差由4.75%降低至2.2%,即横向平移误差平均降低了53.6%;将载体前向平移平均误差由2.2%降低至1.9%,即前向平移误差平均降低了15.4%,长时间运行累积误差率较低,能够满足低功耗高效率计算条件下的组合导航实时载体位姿估计需求。

For solving the optimization problem in vehicle ego-motion estimation in mobile mapping system (MMS) or autonomous vehicle(AV), the relationship between vehicle pose and optical flow is proposed to be utilized and all optical flow vectors are decoupled into 3 translational components, 3 rotational components and1 depth component.The error influences on single component and combined components to vehicle pose estimation are derived.The validity of error separation models with single or combined components is verified through simulation and real-scene data experiments.The combined components error separation model is employed in the proposed flow-decoupled motion field based pose optimization algorithm for stereo visual odometry.Experiment results illustrate that, in conditions of almost the same calculation efficiency as the initial estimation process, this algorithm can reduce the average lateral direction error from4.75% to 2.2%, which means the lateral direction error is reduced by 53.6%;and it can reduce the average forward direction error from 2.2% to1.9%, which means the forward direction error is reduced by15.4%.The results demonstrate that the lower cumulative error ratio can satisfy the requirement of real-time vehicle ego-motion estimation in low power dissipation and high efficiency situations in integrated navigation.