基于在线支持向量机的Mean Shift彩色图像跟踪

New mean shift tracking for color image based on online support vector machine

为了解决传统Mean Shift跟踪方法中目标模板只能从单一图像建立,且很难更新问题,提出了一种新的Mean Shift彩色图像跟踪方法。将RGB颜色空间投影到HSV颜色空间,建立了基于HSV颜色空间的统一直方图核函数模型。为了实现模板在线更新,引入在线支持向量机,推理了基于HSV空间的在线支持向量机的Mean Shift跟踪算法,从而适应目标因尺寸、姿态及光照造成的模型变化。为了验证算法的有效性,对两组国际通用的CAVIAR彩色图像序列进行了跟踪测试。实验结果表明,提出的改进算法在目标姿态、光照或背景发生较大变化时,能有效跟踪目标。当图像分辨率为384×288(目标尺寸约为20×80)时,最快处理速度达40f/s,且跟踪精度比传统Mean Shift提高32.1%。

In order to solve the problem that the object template of traditional Mean Shift tracking can only be built from a single image, and difficult to update, an improved Mean Shift tracking algorithm for color image is proposed. Firstly, RGB color space is projected to HSV color space, and a unified histogram kernel model based on HSV color space is set up. Secondly, in order to achieve template online update, an online support vector machine is introduced, and a Mean Shift tracking algorithm integrated with online SVM based on HSV color space is reasoned. By above operation, object modeling is adaptive to object size, posture or illumination changes. Finally, tracking test on two groups of international general CAVIAR color image sequence is taken to verify effectiveness of the algorithm. Experiments show that the improved algorithm performs well with great changes taken in target pose, illumination or background. When the image resolution is 384 pixel×288 pixel（target size of about 20 pixel×80 pixel）,the fastest processing speed reaches 40 f/s, as well as tracking precision increases by 32.1% than traditional Mean Shift.