Islamic Azad University (IAU), South Tehran Branch, Tehran, Iran.
ISA Trans. 2012 May;51(3):485-97. doi: 10.1016/j.isatra.2012.02.002. Epub 2012 Mar 10.
Contemporary research is developing techniques to tracking objects in videos using color features, and the mean shift (MS) algorithm is one of the best. This known algorithm is employed to find the location of an object, in image sequence, by using a coefficient called the Bhattacharyya coefficient. This coefficient is calculated through an object tracking algorithm to present the similarity in appearance between an object and its candidate model, where the best representation of an object is acquired, once this is could be maximized. However, the MS algorithm performance is confounded by color clutter in background, various illuminations, occlusion types and other related limitations. Because of such effects, the algorithm necessarily decreases the value of the Bhattacharyya coefficient, indicating reduced certainty in the object tracking. In the present research, an improved convex kernel function is proposed to overcome the partial occlusion. Afterwards, in order to improve the MS algorithm against the low saturation and also sudden light, changes are made from motion information of the desired sequence. By using both the color feature and the motion information simultaneously, the capability of the MS algorithm is correspondingly increased, in the present approach. Moreover, by assuming a constant speed for the object, a robust estimator, i.e., the Kalman filter, is realized to solve the full occlusion problem. At the end, experimental results on various videos verify that the proposed method has an optimum performance in real-time object tracking, while the result of the original MS algorithm may be unsatisfied.
当代研究正在开发使用颜色特征跟踪视频中目标的技术，均值漂移（MS）算法是其中最好的之一。该已知算法用于通过使用称为巴氏系数的系数在图像序列中找到目标的位置。该系数通过对象跟踪算法来计算，以表示对象与其候选模型之间的外观相似性，其中获得对象的最佳表示形式，一旦可以最大化该系数。但是，MS 算法的性能受到背景颜色混乱，各种光照，遮挡类型和其他相关限制的影响。由于这些影响，该算法必然会降低巴氏系数的值，表明对象跟踪的确定性降低。在本研究中，提出了一种改进的凸核函数来克服部分遮挡。之后，为了提高 MS 算法对低饱和度和突然变亮的性能，从期望序列的运动信息中进行了更改。通过同时使用颜色特征和运动信息，相应地提高了 MS 算法的能力。此外，通过假设对象的速度恒定，实现了鲁棒估计器，即卡尔曼滤波器，以解决完全遮挡的问题。最后，在各种视频上的实验结果验证了所提出的方法在实时目标跟踪方面具有最佳性能，而原始 MS 算法的结果可能并不令人满意。
