摘要： 针对机器人视觉稳像问题，建立六参数仿射图像运动模型，给出其递推关系。设计基于梯度的KLT特征提取算法，根据最优绝对误差和进行特征点的匹配，利用超定的运动参数求解方程推导，得到有意运动参数的观测模型，并使用最小二乘法进行求解，对卡尔曼滤波后的运动参数和图像运动模型进行反向求解，实现含抖动视频的稳像补偿。在自主移动机器人平台上的实验结果表明，利用KLT算法得到的特征点分布更合理，速度更快，经相对参数滤波后的图像相比绝对参数滤波更平滑。

关键词: KLT算法, 特征匹配, 运动模型, 卡尔曼滤波, 稳像

Abstract: Video stabilization is the key of robot vision. This paper establishes an image affine kinematics model with 6 parameters and its recurrence relations. Kanade-Lucas-Tomasi(KLT) feature matching method is designed based on grads. Optimization of sum of absolute difference is used to match feature points. Through analysis of over-determined image motion equations, observation model of intended motion parameters is derived, and the least squares algorithm is used to solve equations. Through reverse computing of kinematics model using filtered parameters, jitter is compensated and stabilized images are achieved. Experiments on autonomous mobile robot test-bed show that the feature points are uniform distributed and matching is faster by using KLT algorithm with sub-regional fast computing, and the relative parameters filter effect is smoother than the absolute parameters filtering.

Key words: Kanade-Lucas-Tomasi(KLT) algorithm, feature matching, motion model, Kalman filtering, image stabilization

中图分类号: 

• TP242