为了在动态场景图像序列中准确地完成全局运动估计, 实现对运动背景的补偿, 提出了基于均值聚类和几何关系的运动背景估计算法。首先, 利用Harris算法提取两帧图像的特征点, 建立特征点匹配对。其次, 利用K-means聚类算法去除在匹配过程中存在的明显错误的特征点对。再次, 利用三角几何关系去除位于运动目标上的特征点。最后, 利用随机样本一致(RANdom SAmple Consensus, RANSAC)算法和最小二乘方法求出运动参数。分析实验结果得出: 本文算法比原始算法的峰值信噪比提高了5%左右, 所耗时间减少了50 ms。实验结果表明: 该算法能更加精确的实现运动背景估计, 提高了运动背景估计的鲁棒性, 同时提高了计算速度。

A new algorithm for moving background estimation based on means clustering and triangulation is proposed to exactly obtain global motion estimation in dynamic scene and to realize moving background compensation. Firstly, Harris algorithm is used to extract feature points of two frames and initialize the feature point matching pairs. Secondly, K-means clustering algorithm is used to remove the apparent error feature point matching pairs. Thirdly, triangulation is used to remove the feature points in moving target. Lastly, RANdom SAmple Consensus algorithm and least square method are used to solve moving parameters. With analysis on the results, it is concluded that the PSNR of our algorithm is about 5% larger than original algorithm and the time of our algorithm used is 50ms lesser than original algorithm. The results indicate that moving background estimation can be realized more precisely and the robustness of moving background estimation is improved by our algorithm. Furthermore, the computing speed is also raised by the algorithm.