频域与空域处理相结合的大运动偏移稳像算法及其实时实现

Image stabilization algorithm based on processing in frequency-space domain for large motion offset and its real-time implementation

提出了一种频域极坐标处理与空域高斯-牛顿最小化法相结合的运动估计算法。首先使用频域极坐标处理粗估计出两帧图像间的偏移量,将此偏移量作为高斯-牛顿最小化法的初始值,再使用高斯-牛顿最小化法精确估计图像间的偏移量,并补偿图像得到稳定的图像序列。最后,为使算法能够实时运行,针对TMS320C6416数字信号处理器的运行环境优化了算法实现代码。实验结果表明,优化后的算法能够实时、高精度地稳定大运动偏移图像序列。

A motion estimation method combining the polar coordinates transform in the frequency domain with the Gauss-Newton minimization in space domain is presented. Firstly, a rough offset value between two frames＇ images is coarsely estimated using the polar coordinates transform in the frequency domain. Then the accurate offset between two frames＇ images is efficiently estimated using Gauss-Newton minimization method by taking the rough offset as the initial value. Therefore, a stable video sequence by image compensation using the accurate offset is achieved. Finally, the algorithm is optimized according to the environment of TMS320C6416 digital signal processor （DSP） for real time processing. Experimental results show that the optimized algorithm can accurately stabilize the video sequence with large motion offset in real time.