用于Android手机活动识别的深度重构模型

Deep reconstruction models for activity recognition using Android Phones

基于数据的流形假设,进行了Android手机的活动识别研究,提出了一种深度重构模型(DRMs),该模型无需基础几何的先验假设就能自动学习到当前类样本的复杂非线性曲面结构和几何特点。首先定义了一个深度重构模型(DRM)模板,通过高斯受限玻尔兹曼机(GRBMs)逐层贪婪训练以初始化DRM模板的权重。在训练阶段,根据每类样本分别微调初始化后的DRM模板便可得到特定类的DRM。在测试阶段,基于测试样本与特定类DRM的最小重构误差决策活动类别。实验证明,该方法对Android手机数据集的活动识别正确率高达99%。

The activity recognition using Android phones is studied based on the assumption of the manifold-shaped data,and a deep reconstruction model（ DRM） which can learn the complex nonlinear curved surface structure and geometric features of current class samples without a priori assumption of basic geometry is proposed. Firstly,a template of the DRM is defined,and its parameters are initialized by performing the unsupervised pre-training in a layer-wise fashion using Gaussian restricted Boltzmann machines. In the training stage,the initialized DRM template is then separately trained for training the samples of each class and the class-specific DRMs are learnt. In the testing stage,activities are recognized based on the minimum reconstruction error between the learnt class-specific models and the test samples. The experiment performed using the Android mobile phone dataset show that the correct rate of this method for activity recognition is up to 99%.