基于手机传感器的握持方式判断及运动状态识别

摘要

Traditional activity recognition methods are based on sensors at the fixed positions of users. Once the sensors' positions are changed, the performance of the methods will be degraded. Unlike most of these studies, the proposed system firstly detects the holding mode of the phone, and then recognizes the human activities. Our work contains preprocessing, feature extraction and classification. By using wavelet transform and singular value decomposition to extract and reduce features dimension and using RBF-based SVM (support vector machine) for classification, the system is able to recognize 5 holding modes (close to body on the side, swing, holding at the front, close to ear) and 5 activities (stationary, slow walking, normal walking, fast walking and running). Comparing with 4 common classifiers, the result shows that the proposed method performs the best and its detection accuracy is about 93% .%传统运动识别技术多以传感器位置固定为前提进行识别,但当传感器放置位置或握持方式发生变化时运动识别率会受到相应影响.该文提出了一种基于手机传感器的握持方式判断及运动状态识别方法,解决了传感器随放置位置不同影响运动识别率的缺点.该方法首先通过传感器对设备握持方式进行判断,使用不同握持方式下的三轴加速度数据进行特征提取,通过多层小波变换得到各层高频和低频部分,对其进行组合形成初级特征,用奇异值分解对初级特征进行降维得到最终特征,使用基于径向基核函数的多分类支持向量机(SVM)对特征分类,进而判断不同握持方式下的不同运动.实验结果表明,该方法对不同运动方式下的平均识别率为93%.