基于可穿戴传感器的行为识别随机逼近模型

为了使分类器能够在某个强度级别的行为样本集上训练而在其他强度级别上正确分类行为,提出了行为识别的随机逼近模型。在训练阶段从加速度计的时间序列数据提取特征,然后将特征送入聚类算法。数据依据行为聚类,聚类的均值和方差组合成相对应的SAM。在识别随机行为阶段,测试样本和每种行为类别的SAM进行比较。利用聚类算法和随机逼近给每种行为创建模型,然后使用启发式随机逼近最近邻方法来对行为进行分类。在实验中结合k-均值和高斯混合模型两种聚类算法,验证了提出的随机逼近模型的性能优于其他几种流行的行为分类方案。

Activity Recognition Using Stochastic Approximation Model Based on Wearable Sensor

In order to make classifiers can train on an activity at a subset of intensity levels and classify the same activity at other intensity levels,an activity recognition using stochastic approximation model based on wearable sensor is proposed in this paper. The training phase begins by extracting features from the accelerometer time series data,then put the feature into a clustering based algorithm. The data is grouped by activity into clusters,and the clusters＇ mean and variance are combined to form a corresponding SAM. In the recognition phase,the sample test points are compared against each activity category ＇s SAM. Clustering algorithms and stochastic approximation are utilized to create a model for each activity,and then use a stochastic approximation nearest- neighbor heuristic for activity classification.Experiments are reported for each dataset using two clustering algorithms,k- means and Gaussian mixture model. The stochastic approximation model is superior to other popular activity classification schemes.