为了实现对人类行为的有效识别,提出了一种基于层次聚类多任务学习(HC-MTL)的人类行为识别方法.采用正则化最小二乘法制定目标函数,并对模型参数和分组信息这2个潜在的变量进行联合优化.使用聚类范数正则化方式进行多任务学习,并求解任务相关性,进而对人类行为进行有效识别.该方法打破了所有行为是独立的个人学习的假设,通过任务聚类的方式建立起多任务之间的关系,共享同类任务之间的相关信息,提高了人类行为识别的准确度.试验结果表明,与聚类多任务学习方法(CMTL)和鲁棒多任务学习方法(RMTL)相比,HC-MTL方法可以发现任务的潜在相关性,有助于诱导群体多任务学习.通过同一类任务之间的共享信息,减少了误差,并提高了行为识别的精确度.%In order to realize the effective recognition of human behavior,the method based on hierarchical clustering multi-task learning(HC-MTL) is proposed. Using the regularized least square method,the objective function is formulated,and the two potential variables of model parameter and grouping information are jointly optimized. The multi-task learning is conducted and the task relevance is solved with the clustering norm regularization approach,then the human behavior is effectively recognized. The method breaks the hypothesis, i. e. , all the behaviors are independent individual learning, and through the way of task clustering,the relationship between multi-task and the relevant information among similar sharing tasks are established,thus the accuracy of human behavior recognition is enhanced. The test results show that compared with clustered multi -task learning method( CMTL) and robust multitasking learning method( RMTL) ,the HC-MTL method can discover the potential relevance of the tasks and help to induce the group multi-task learning,through sharing information among the same types of tasks,the error is reduced and the accuracy of behavior recognition is improved.
展开▼
