车辆识别中的博弈特征选择研究

针对特征识别判断中出现的冲突和特征过多所造成的特征维数灾难问题,采用博弈论的方法对车辆特征空间进行搜索,找到一组对车辆类型识别的最佳特征子集。在多特征组成的特征空间中把具有存在冗余关系的特征定义为博弈过程中的同一个局中人,具有互补关系的特征定义为不同的局中人,用互信息构造支付值函数,使支付值函数能够表现局中人之间的可信度冲突关系。支付值函数的大小和局中人采用对策的分类能力大小相关,分类可信度高的对策支付值大,分类可信度弱的对策支付值小。在车辆识别应用中,车辆的5个特征所组成的特征空间中,把具有冗余与关系的特征分成两个局中人,构造出支付函数,在赢得矩阵中求出均衡解,得出均衡解所对应的{车长、后轮距}是特征空间的一组最佳特征子集。

Research of Feature Selection for Vehicles Recognition Based on Game Theory

In order to solve the problem of characteristic dimension disaster caused by excessive features and conflicts in the identification and judgment of characteristics,a subset of the best features for vehicle identification was searched out in vehicle feature space using game theory.The features which have redundant characteristics were defined as the same player,the features which have characteristics of complementary were defined as different player.Payoff function which can reflect the conflict of credibility between players was constructed using mutual information.The value of payoff function is relevant to the classification capability of player.The payoff value of countermeasure which has high classification reliability is high,and vice versa.In vehicle identification,the features which have redundant characteristics were found out as two different players in the space of five vehicle features to construct payoff function.The equilibrium solution was obtained from the payoff matrix,and its corresponding {vehicle length,rear gauge} is the subset of the best features of feature space.