摘要
为解决视觉跟踪中线性RGB组合特征的权值仅使用整数这一问题,设计了一种将权值范围扩大到实数域的特征及其选择算法。在线性RGB组合特征的权值向量空间中,能够最有效地区分前景和背景的权值向量就是能够用来跟踪的最佳权值向量。将线性RGB的组合扩展至RGB的二次组合,使用逻辑回归对前景和背景像素进行分类,并利用加权迭代最小二乘法对逻辑回归求得的权值向量进行迭代,得到最有效区分前景和背景的权值向量,使用特征更新策略对目标权值向量进行不断更新。实验结果表明,改进的方法结合Means Shift可有效的跟踪运动目标。
To enlarge the weight of linear combination of RGB features from only integers to the field of real numbers, an improved discriminative RGB feature space and related feature selection method are presented. Based on experience, in the weighted RGB feature space, the features that best discriminate between object and background are also the best for tracking the object. First, the linear combination of RGB is enlarged to quadratic combination of RGB. Then, the logistic regression is applied to classify the object and the background pixel, and the iteratively re-weighted least squares (IRLS) is used to ensure that the method can discriminate between object and background efficiently. At last, feature update method ensures the tracking process proceeds continuously. Experiment results demonstrate that this method combined with the Mean Shift tracking algorithm can figure out the object more effectively.
出处
《计算机工程与设计》
CSCD
北大核心
2013年第12期4284-4288,共5页
Computer Engineering and Design
关键词
加权RGB特征
目标跟踪
逻辑回归
最小二乘法
特征更新
weighted RGB feature
target tracking
logistic regression
iteratively re-weighted least squares
feature update