Research Of Human Action Recognition Based On Combination Of Multi-sensor

There are two aspects of Human Action Recognition,human action recognition based on analysis and human action recognition based on sensors.Among the two ways mentioned above,sensor based human action recognition has been widely applied to fields of smart bracelet and wearable body sensor network because of its advantages that by witch it could directly measure the motion data during human activity,and that it could hardly to infringe on personal privacy.It also has uncomplicated algorithm at the same time.With the development of sensor technology and the continuous research of machine learning algorithm,this kind of research is developing towards the diversification of sensors,the diversification of data and the diversification of sensors combination.Sports is also a field in which sensor based human action recognition could play a role in.Obtaining the human data in motion can acquire the characteristics of human movement and be helpful to the improvement of sports level.Data is provided by sensor for technical motion analysis.In this study,some human daily movement action and some technical motions of basketball are involved.The main research of this paper is to identify human actions by multi-sensor.The target nine kinds of actions are standing,walking,running,free throwing,jumping shooting,walking dribbling and running dribbling.The contents of this paper include the combination of the number and position of the sensor,wavelet threshold denoising,recognition of the decision tree and the SVM classifier.In this research,analysis of the target recognition,the design and construction of the experimental system,the environment and experimental design of the experiment,and the evaluation of the results are also involved.This paper presents a recognition algorithm to identify the characteristics of weight loss,this algorithm,which based on vertical acceleration analysis,has accuracy rate of 98.67%.To be trained,this algorithm need inputs of estimation values of jumping height.In this paper,300 groups of data are collected of 10 persons by two ways of sensor location combinations.Through the comparation of two different sensor location combinations,this paper has drawn three conclusions.(1)These two kinds of combinations both can achieve high recognition rates(96.49% and 96.20% using decision tree,99.40% and 98.97% using SVM);(2)The recognition rates of these two ways of sensor locations are likely to be equivalent;(3)SVM classifiers is better than decision trees to recognize target actions in this paper;(4)The weight loss detecting algorithm are effective considering both of its accuracy and its time cost.