摘要
隐马尔科夫模型(HMM)是人体行为识别领域使用最普遍的理论方法,但传统的HMM训练方法Baum-Welch(BW)法是一种爬山算法,容易陷入局部最优解,从而影响行为识别的准确率。针对这一问题,本文采用遗传算法来训练HMM的参数,首先对HMM的主要参数状态转移概率矩阵A和输出分布矩阵B进行实数编码,用对数似然概率评价个体适应度,最后进行遗传运算。实验结果证明GA算法能够较好的克服传统HMM训练容易陷入局部最优的问题,GA-HMM比BW-HMM在人体行为识别时有更高的识别率。
Hidden Markov Model（HMM） is the most commonly used method in the field of human behavior recognition .The traditional training method Baum-Welch（BW） is a kind of hill-climbing algorithm, which is easy to fall into local optimal solution, affecting the recognition rate. To solve this problem, we ues genetic algorithm to train the HMM parameters. In the first, we make the main parameters of the HMM state transition probability matrix A and output distribution matrix B real-coded, then use logarithmic likelihood to evaluate individual fitness and take the genetic operator in last. Experimental results show that the GA is able to overcome the problem that traditional HMM training is easy to fall into local optimum, GA-HMM has a better performance than BW-HMM in human action recognition.
出处
《电子测试》
2013年第2期88-90,共3页
Electronic Test
