Abstract: Recently, fusion-based gait recognition has become one of the hottest topics in the domain of biometrics recognition. There are not only complementary relation among different information sources, but also redundancy and conflict. A novel fusion method based on game selection model is proposed from the feature selection standpoint. Firstly, the features which have redundant characteristics are defined as the same player, and the features which have characteristics of complementary are defined as the other player. Then payoff function is constructed using entropy and Mutual Information（MI）, which can reflect the conflict of credibility between players. At the same time, the payoff value reflects the distinguishing power of the classification. The higher the payoff value is, the stronger the classification ability of features are. Finally, the equilibrium solution is obtained from the payoff matrix by maximin principle, and its corresponding position is the optimal subset. Through the exhaustive comparison experiments on CMU and CASIA data sets, the effectiveness of the proposed method is verified. Experimental results show that the best performance of recognition can be obtained by using less computational complexity.

Key words: game theory, feature selection, fusion-based gait recognition, payoff function

摘要： 基于信息融合理论的步态识别已成为生物特征识别领域最为活跃的研究方向之一。不同信源之间不仅有互补关系，也存在冗余和冲突。从特征选择的角度对这一问题进行研究，提出了基于博弈选择模型的融合方法。根据特征的冗余和冲突关系定义局中人和策略集；用信息熵和互信息构建支付函数，并计算出模型的支付矩阵；再利用极大化极小原理求解博弈均衡点，从而得到最佳策略组合。使用模型选出的最佳策略进行步态融合，以得到较高的识别和校验性能。通过在CMU和CASIA数据集上的穷举组合对比实验，验证了所提方法的有效性，并且该方法可在运算量较小的情况下取得最佳识别性能。

关键词: 博弈论, 特征选择, 步态融合, 支付函数