提高Mean-shift跟踪算法性能的方法

针对Mean-shift跟踪算法,在目标色彩特征不突出,或受到光照、阴影等影响,或有干扰物体靠近目标时,很容易发生跟踪错误等问题,采用色彩融合模版和位置预测来提高Mean-shift跟踪算法的性能.在对图像的RGB三色进行简单的线性融合的基础上,提出了根据前景和背景直方图的相似度函数去选取目标特征最突出的融合图像的算法,并据此建立3个目标模版.对目标的位置先进行卡尔曼预测,再用Mean-shift算法对3个模板分别进行跟踪,最后融合跟踪结果.实验结果证明,提出的方法能在复杂背景下跟踪目标,并能更好地应付阴影、光线等变化.此外,它能有效地避免相似物体靠近目标或者和目标交错引起的跟踪失败.

An Approach to Improve the Performance of Mean-shift Tracking Algorithm

It is well known that the Mean-shift tracking algorithm fails in tracking targets when suffering from changes of light,shadows,or/and similar objects approaching to the tracked one.The color fusion of an image and the prediction of the position of tracked moving target are used to enhance the performance of the Mean-shift tracking algorithm.Based on the RGB color fusion of an initial image,with simple linear combination,according to a similarity function between the object and background histograms of the image,an algorithm which can give prominence to the fused features of the image is presented.With the presented algorithm,three templets for a tracked target are selected.After the position of the moving target is predicted by Kalman filter,the Mean-shift algorithm employs each of the three selected templets to track the same tar-get.The final tracking result is a version of the fusion of all three independent tracking results of the three templets.A number of experimental results show that the proposed method can track an object in compli-cated background with shadows and changes of light.Especially,it can effectively avoid the failures due to some similar objects approaching to the target.