Abstract: In the study of radio frequency based indoor passive localization, there may be so many link features that the complexity of classifiers is high. To solve this problem, an indoor passive localization with class-based feature selection method is proposed. To convert a multi-class classification to multiple two-class classifications and use max-min Markov blanket algorithm to select discriminative features for every class and train for related two-class classifiers. In the testing phase, to predict test samples’ class by probability estimates output of support vector machine. The experimental results show that the proposed method not only significantly reduces features and lower the complexity of classifiers, but also improves the localization accuracy.

Key words: indoor passive localization, feature selection, Markov blanket, support vector machine

摘要： 针对基于射频的室内被动定位研究中，用于训练定位分类器的链路特征数较多，分类器复杂度较高的问题，提出了一种基于类别的室内被动定位特征选择方法。此方法将被动定位的多分类问题转化为多个二分类问题，利用最大-最小马尔科夫毯为每个位置类别选择其表征能力较强的特征子集，并构建相应的二分类模型。在测试阶段，采用支持向量机的概率评估输出，最终确定测试样例的位置类别。实验结果表明，此方法大幅减少了定位所需的特征维数，降低了分类器复杂度，同时使定位的准确度得以提高。

关键词: 室内被动定位, 特征选择, 马尔科夫毯, 支持向量机