【摘要】 针对人体行为数据的识别与分类问题,提出一种连续时变自编码机(Continuous Time-varying Autoencoder,CTAE)模型.该模型在激活函数中增加高斯随机单元,强化对非线性连续型数据的特征学习与提取.在人体行为识别实验中,从原始数据信号中提取十维频域特征和四维时域特征;利用主成分分析(Principle Component Analysis,PCA)方法实现特征数据降维;针对预处理完的人体行为数据,训练由多个CTAE组成的深度信念网络(Deep Belief Network,DBN),实现行为识别与非线性分类.仿真验证了模型的有效性.更多还原

【Abstract】 This paper proposed a model of continuous time-varying autoencoder(CTAE)to solve the problem of human behavior recognition and classification.The stochastic unit following the Gaussian distribution was added to the activation function of CTAE,which could change the direction of gradient and prevent over-fitting.In the experiment of human behavior recognition,the ten-dimensional frequency domain feature and four-dimensional time domain feature were extracted from the raw signal.Then,the principle component analysis(PCA)was adopted to reduce the dimension of the feature.After data preprocessing,the deep belief network(DBN)composed of multiple layers of CTAE was trained to realize the nonlinear classification and recognize the human behavior.The effectiveness of the CTAE model was validated by simulation.更多还原