The Research Of Human Action Recognition Method Based On Wearable Motion Capture System

Human behavior is the main expression form of the interaction between a people and the environment,so the researches on the methods by researchers for exploring the human action recognition have not been stopped.Human action data could be acquired from the motion capture system,it is based on wearable sensors and can record human body physical information of each part such as acceleration and angular speed,and it has extremely important meaning in analyzing the essence of human action.So the work on human action recognition based on this system has been a hot spot in the research field.Among the traditional behavior recognition algorithms,the feature extraction is based more on redundancy statistical feature vector space,which not only decreases the efficiency of the algorithm,but also because of the differences of the natural behavior characteristic affects the recognition efficiency of the algorithm.Moreover,the traditional algorithms are carried out with the need of the long time samples which include several periods for the stability of the time and frequency domain features.But in the condition of short time samples,the time and frequency domain features which extracted by traditional algorithms are unstable and will influence the results in a detrimental fashion.This paper focuses on the study of the shortage of the traditional algorithms.First,the traditional algorithms extract high-dimensional features,which lead to the increase of computational complexity.In this view,the paper proposes a new algorithm based on statistical characteristic vector and natural behavior characteristic.We only select 11 features to describethe actions which are based on analyzing the essence of behavior.It effectively compresses the characteristic space and expresses the range and acuteness of body relative to the deflection of the vertical and horizontal direction.Then we use the support vector machine as the classifier and build the classification model.The experimental results show that this presented algorithm can effectively identify 13 kind of daily behaviors of the WARD1.0 database.Second,in the condition of short time sample,the instability of traditional time and frequency domain features will affect the classifier hurtfully.Thus this paper proposes a new algorithm for the short time samples based on template matching.The main idea is making the training template set complete enough with extracting with sliding window and the pattern of each actions contain several atomic-patterns.It is the guarantee which the best and most similar training template could be found for each test template,and we judge the category of test template according to the label of the appropriate training template.Our experimental results show that the algorithm based on template matching acquires good recognition efficiency in the conditional of short time samples.