基于SVM与Mean-Shift的非刚性目标跟踪框架

Nonrigid target tracking framework based on SVM and Mean-Shift

针对动态背景下,序列图像中的非刚性目标跟踪问题,提出了一种基于支持向量机(support vector machine,SVM)和均值移动(Mean-Shift)的序列图像目标跟踪框架。在初始图像中选择跟踪目标所处的矩形框,将目标框周围一定范围的像素作为背景。以目标和背景数据训练SVM二值分类器。运用得到的分类器对下一帧图像相同区域内的像素分类,得到二值的置信图(confidence map),在置信图范围内运用Mean-Shift算法求得当前目标位置,移动目标框和背景框的中心到目标位置,以10%的比例缩放目标框并选择最优者以适应目标尺度变化。以此时的目标像素和背景像素训练新的SVM分类器,进行下一幅图像的跟踪,直至完成整个序列图像跟踪任务。实验证明,该方法适用于动态背景及非刚性目标的跟踪,且实时性较好。

To solve the problem of tracking nonrigid targets for image sequence in dynamic scene, a tracking framework based on support vector maehine（SVM） and Mean-Shift is proposed. The rectangle in which the tracked target lies in is selected in the intial image, and the pixels in the given region around the target are taken as scene data. Then both the target and the scene pixel data are applied to train a two-value SVM classifier. The obtained SVM classifier is employed to classify the pixel data lying in the same region of the next image, thus the confidence map including two pixel values is got. The Mean-Shift algorithm is performed to get the position of the current target within the area of the confidence map. Similarly, the new SVM classifiers are trained for tracking the images to be followed until the whole image sequence tracking is completed. The experiment results show that the proposed method is suitable for tracking nonrigid targets and active scenes, and has the real-time characteristic.