Abstract: LDA is one of the most commonly used feature selection methods in human activity recognition based on the acceleration signal. While its objective function is not the training error, it cannot ensure that obtained projective subspace is with the minimum training error. To solve this problem, a modified method that uses LDA based on GA is proposed. Due to the small sample size problem, the original samples should be processed by PCA first. Then, to achieve the minimum training error, the method adjusts eigenvector of the between-class scatter matrix by using GA. 7 activities of daily living are identified by SVM. Results of tests show that compared with using PCA only and using PCA+LDA, LDA based on GA can effectively raise the recognition accuracy. In addition, it can also reduce the feature dimension and decrease the training error. It obtains an average accuracy of 95.96%.

Key words: activity recognition, accelerometer, Principle Component Analysis（PCA）, Linear Discrimination Analysis（LDA）, Genetic Algorithm（GA）, Support Vector Machine（SVM）

摘要： 在基于加速度信号的人体行为识别中，LDA是较常用的特征降维方法之一，然而LDA并不直接以训练误差作为目标函数，无法保证获得训练误差最小的投影空间。针对这一情况，采用基于GA优化的LDA进行特征选择。提取加速度信号特征，利用PCA方法解决“小样本问题”，通过GA调整LDA中类间离散度矩阵的特征值矢量，使获得的投影空间训练误差最小。采用SVM对7种日常行为进行分类。实验结果表明，与单独采用PCA和采用PCA+LDA方法相比，基于GA优化的LDA算法在保证较高识别率的同时能有效降低特征维数并减小分类误差，最终测试样本的识别率可达95.96%。

关键词: 行为识别, 加速度传感器, 主成分分析（PCA）, 线性判别分析（LDA）, 遗传算法（GA）, 支持向量机（SVM）