一种面向高斯差分图的压缩感知目标跟踪算法

针对压缩感知目标跟踪算法在目标纹理改变、比例缩放、光照变化剧烈时鲁棒性不足,提出一种面向高斯差分图的实时跟踪算法.首先,构建图像的多尺度空间及其对应的高斯差分图,实现高斯差分图的特征提取并获取压缩感知的输入信号;然后,通过压缩降维,目标邻域遍历,参数更新等过程,计算出面向高斯差分图的后续帧的目标最优跟踪窗;最后,将跟踪窗投影到对应的原始图像上,完成面向视频流的目标跟踪.高斯差分图像是单通道灰度图,具有灰度取值范围小、数值低、结构简单、维数少等特点,增强了特征对纹理改变、比例缩放和光照变化的稳健性,且继承了传统算法的实时性.实验证明,该算法能够快速准确地实现复杂环境下的移动目标跟踪任务.

Target tracking by compressive sensing based on Gaussian differential graph

As traditional target tracking based on compressive sensing has poor robustness in texture change, scale variation and illumination change, a real-time tracking algorithm using compressing sensing based on Gaussian differential graph was proposed. Firstly, Gaussian differential graph is acquired from multi-scale space of image. The features are extracted from the graph and taken as input signals of impressive sensing. Secondly, by compressing, dimension reduction, target neighborhood traversal, parameters update, the optimal search window is estimated. Thirdly, the search window is mapped onto the corresponding original image, and target tracking in the video sequences is finished. Gaussian differential graph had some characteristics such as single-channel, small grayscale range, low value, simple structure, small dimensions, which make the algorithm have strong robustness in scaling, texture and illumination changing. The real-time performance was inherited from the traditional algorithm. Experiments proved that with the proposed algorithm the moving target can be tracked quickly and accurately in a complex environment.