基于隐动态条件神经域的在线行为识别方法

摘要

In view of the continuous unsegmented human behavior recognition in video,a kind of online behavior recognition algo-rithm based on latent-dynamic conditional neural field (LDCNF)was introduced.LDCNF model contained two hidden layers,on the basis of latent-dynamic conditional random field (LDCRF),a layer of neural network was added,i.e.gating layer,to extract non-linear relationships between input data and output labels.A new regularization term was added for the training of this mo-del,encouraging action sequences’diversity between hidden-states.In the simulation experiment,ten kinds of behavior re-cognition results for conditional random field (CRF),HCRF,LDCRF and LDCNF were compared.For the online processing be-havior sequence,the results show that the proposed algorithm,compared to CRF,HCRF,LDCRF,has better recognition rate.%针对视频中连续的未分割人体动作识别存在的一些问题,提出一种基于隐动态条件神经域模型(latent-dynamic conditional neural fields,LDCNF)的在线行为识别方法。LDCNF模型含有两个隐层,在潜动态条件随机场(LDCRF)的基础上,增加一层神经网络层,即门层,提取输入数据和输出标签间的非线性关系；增加一种新规则项训练该模型,辨别动作序列隐状态间的差异性。在仿真实验中,针对10种连续的行为动作,将该算法与条件随机场(CRF)、HCRF、LD-CRF进行识别效果的对比。实验结果表明,对于联机处理行为序列,该算法相比于 CRF、HCRF、LDCRF模型具有更好的识别率。