一种极速学习机人体行为识别模型迁移方法  

Human activity recognition transfer learning based on extreme learning machine

作　　者：王忠民[1] 屈肃 

机构地区：[1]西安邮电大学计算机学院,陕西西安710121

出　　处：《西安邮电大学学报》2015年第1期49-54,共6页Journal of Xi’an University of Posts and Telecommunications

基　　金：国家自然科学基金资助项目(61373166);工业和信息化部软科学研究计划资助项目(2014R32);陕西省工业攻关计划资助项目(2012K06-05);陕西省教育厅产业化培育基金资助项目(2012JC22)

摘　　要：针对移动用户行为识别的问题,提出一种带权值样本筛选的迁移学习方法。该方法通过将训练集分割重构并赋样本权值,依据训练样本在对应极速学习机分类器上的判别结果对其权值进行修改,经过多次迭代后筛选出与迁移样本最相似的训练集,从而构建迁移行为识别模型。测试结果表明,迁移学习后的行为识别模型能有效提高分类正确率。To increase the accuracy of the mobile user activity recognition, a novel algorithm named Transfer Learning Based on Weighted Sample Selecting （TLWSS） is proposed. By reconstructing the sampling training set segmentation and assigning weights, TLWSS can modify the weights on the basis of the training sample discriminant results in corresponding Extreme Learning Machine （ELM） classifier. It has the ability of screening more similar sample training set with the transfer set, especially after many iterations. Eventually the final activity recognition model will be built. Experimental results show that the activity recognition model after transfer learning has higher classification accuracy.

关 键 词：行为识别 迁移学习 极速学习机 集成学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]