摘要
为研究面向动作捕捉的非线性时间序列预测的方法。通过对人体动作数据进行分析,研究并实现基于动作捕捉数据的预测方法,解决因传感器故障而引起的数据丢失、修正问题。通过模拟实验假设动作序列中某一个传感器发生故障,随后使用8种机器学习方法,利用6种指标进行评估,对比各种方法的预测效果,并将预测后的动作进行可视化。通过实验,随机森林、决策树、最近邻方法对数据的预测准确率能达到90%以上。由此,面向动作捕捉的非线性时间序列预测方法能够准确地还原动作。
In this paper, we study the nonlinear time series prediction method for action capture. A prediction method based on the capture data is studied and implemented by analyzing human motion data to solve the data loss and correction problem caused by sensor failure. Based on this research purpose, the simulation experiment assumes that a sensor in the sequence of actions fails, then uses eight kinds of machine learning methods, and evaluates them with six indexes. The prediction results of different methods are compared and the predicted motions are visualized. Through the experiments, data prediction accuracy by random forest, decision tree, nearest neighbor（KNN） method can reach more than 90%. Thus, the nonlinear time series prediction method for motion capture can accurately reconstruct the action.
作者
黄天羽
郭芸莹
Tianyu Huang;Yunying Guo(Beijing Institute of Technology, Beijing 100081, China)
出处
《系统仿真学报》
CAS
CSCD
北大核心
2018年第7期2808-2815,共8页
Journal of System Simulation
关键词
动作捕捉
非线性时间序列预测
机器学习
性能评估
动作预测
motion capture
nonlinear time series prediction
machine learning
performance evaluation
action prediction
