改进的单目视觉实时定位与测图方法

针对经典单目实时定位与测图(SLAM)采用卡尔曼滤波(EKF)滤波和FAST特征角点所存在的非线性误差和鲁棒性较差的问题,提出了一种改进的单目视觉实时定位与测图方法。该方法采用相机中心的迭代EKF(IEKF)滤波方法,将特征点在当前相机坐标系下表达,并在线性化展开点附近迭代更新,不断逼近最优位置,从而最小化线性化误差;针对特征点跟踪的鲁棒性、高效性及分布不均的问题,选用具有尺度和旋转不变性,且探测和匹配效率更高的ORB特征作为特征角点,并采用一种由探测到筛选阶段的整体网格化处理方法;另外,采用特征点逆深度参数化方法,避免了因深度信息未知而导致的局部地图初始化错误问题,并采用1点随机抽样一致方法(RANSAC)滤波更新方法剔除错误的特征匹配,保证滤波估计的准确与稳定。实验采用外符合精度对算法进行评价,结果表明:新方法具有更强的鲁棒性,绝对定位精度提升至2.24 m,误差轨迹比提升至1.3%,且满足实时性要求,是一种实用性较强的单目视觉实时定位与测图方法。

Improved monocular simultaneous localization and mapping solution

An improved monocular simultaneous localization and mapping method is proposed to solve a series of problems existing in the conventional monocular SLAM system, which is based on the classical EKF filter and FAST corners. To reduce the state estimation error resulting from deviation of expansion point when linearization, the camera centered iterated EKF is applied to monocular SLAM system, which can minimize linearized error by iterative updating and representing all feature locations in the current camera frame. For robustness and efficiency of tracking features, and a homogeneous distribution of feature points, ORB features, which have the property of fast detection and matching, and invariance to scale and rotation, are selected as the feature points. Moreover, cell division method through detection to selection is applied. And the utilization of inverse depth parameterization for point features can efficiently avoid the problem of scalable ambiguity when map initialization. In addition, 1 point RANSAC approach can ensure the stability and precision of filter by eliminating the wrong feature matching. The performance of system is evaluated by the ground truth. Experiments show that this new method is more robust and precise in comparison with the other monocular SLAM solutions, and meets the real time processing requirements. The absolute trajectory positioning precision increases to 2.24 m and the mean error over the trajectory increases to 1.3%. To sum up, the proposed method is a practical solution to monocular localization and mapping.