Research On Skeleton-Based Human Action Recognition Method

Human action recognition is a popular topic in computer vision and pattern recognition and plays a pivotal role in the "human-centered computing".Most traditional researches focus on the video images.Under the era of big data with the rapid development of machine learning technology,the study on motion data via sensors has become popular in recent years.More attention will be paid on the data-driven human action recognition inevitably.Where,the skeleton-based human action recognition has become a hot-spot issue.Feature extraction and classifier construction are key steps in the classification algorithm and the core issue in human action recognition.Most researches use single-level classification,which ignores the internal hierarchical relationships.About features,most studies pay less attention on time issues and transform motion time series into disordered data.However,time is also an important motion feature.Therefore,we give a hierarchical classification method,not only consider action gradation,but also the temporal and spatial features.Furthermore,we design an optimal classification unit to improve the classifier construction,which can automatically select an optimal classifier by data dependence,and enhance recognition effect.Specifically,we propose a method based on hierarchical recognition strategy in Section 3.According to human anatomy and human kinematics,with the action granularity,we extract different features and select the proper classifier in each level.In Section 4,we think deeply and develop the self-selection classifier method.The optimal classification unit is designed for eliminating the influence of subjective factors and experience,and select most suitable classifier by training each classifier with data.Probability voting mechanism is used to ensure the reliability of the choice.The proposed methods perform excellent on public and self-built datasets.