基于Bayes规则与HMM相结合的步态识别方法研究

提出一种将Bayes规则与HMM相结合的步态识别方法.检测环节通过采用重心变化作为特征序列来削弱时间差分算法中运动实体目标存在中空的负面影响,通过对检测出的目标特征序列采用直线拟合提取对称轴,并等效转化为具有方向性的点线距序列,来简化运算,降低失真度.步态训练中,通过初始化的修正使训练出的HMM接近全局最优,并给出Bayes相关先验学习方法.步态识别中,应用HMM的前-后向算法,并融合Bayes规则,客观性增强,最终结果经中国科学院CASIA的原始步态视频测试,达到比较高的识别率,且对衣着具有一定鲁棒性.该文方法主要针对视角在0°～180°间的室内监控直道行走场景.

A Method of Combining Bayes Rule with HMM in Gait Recognition

This paper presents a framework of combining Bayes rule with Hidden Markov Model(HMM) to recognize human identification by gait,indoors.First,the monitored human motion is detected mainly by a three-frame differencing algorithm.Then,a curve of centroid on the object’s motion can be acquired.The curve is transformed into the observation sequence of its corresponding HMM by adaptive filtering,median filtering,line fitting,rotating equivalently,normalizing,nearest neighbor clustering,and cycle extracting in turn.During the process of training the HMM with Baum-Welch algorithm,the original parameters of matrix B is modified statistically by Viterbi algorithm,making the final trained model approximate global optimization further.And the prior knowledge in the Bayes rule is also acquired from relative learning.Lastly,the observation sequence is used to recognize the human identification by means of combining Bayes rule with the Forward-Backward algorithm in trained HMM.In the end,the performance of the method is illustrated by the videos of the CASIA Gait Database,the result acquires comparatively higher recognition rate and is robust for the objects’ clothes to a certain extent.The framework of this paper is fitting for monitoring indoors.The type of the gallery monitored is straight and the sight angle for object is between 0° and 180°.