Study On Human Posture Recognition Based On Mechine Learning

With the development of technologies such as the Internet of Things,gesture recognition plays an increasingly important role in important fields such as biomedicine,human-computer interaction,and smart home.my country's population aging trend is intensifying,and the elderly population is increasing,which directly or indirectly causes a large amount of social burden and economic loss.Human body gesture recognition technology plays an important role in smart home care and remote medical technology.In this paper,research on human pose recognition based on portable sensor system is carried out,mainly based on the instantaneous and time domain characteristics of human pose sensor data,two recognition algorithms are constructed respectively,and the algorithm is verified in the physical system.The human pose recognition system has the problems of unstable sampling data and high system cost.On the one hand,the poor combination of software and hardware of the human posture data acquisition system leads to large acquisition errors.On the other hand,current pose recognition algorithms have problems such as poor robustness,huge time cost,low recognition performance,and few types of recognition.In this paper,the main research contents are as follows:(1)Design a machine learning algorithm based on the instantaneous characteristics of human posture.The data acquisition system collects 22-dimensional acceleration signals,angle signals,and plantar pressure signals from key points of the human body,and further designs a fusion feature combining multidimensional sensors.Considering the portability and robustness of the algorithm,a random forest algorithm is used for pose recognition.Further combining the data discretization method of clustering thought,the data discretization of the random forest algorithm is performed The effects of the combination of a variety of clustering algorithms with different forests and random forests are compared,and the distance parameters of the sensitive parameters in the clustering algorithm are compared.Finally,the residual distance is determined as the best distance indicator.DBSCAN-RF pose recognition algorithm.The inheritance idea of the Boosting algorithm is studied,and the Adaboosting classifier is improved in combination with the cascade idea,which further reduces the false positive rate of the human pose recognition algorithm and improves the recall rate of the algorithm(2)Realization of Markov model modeling of various poses of human body.The eight common poses of the human body are regarded as a continuous process,and their Markov properties are used for modeling.Algorithms based on instantaneous features have a very high rate of false positives for retained transition data,and it is impossible to focus on continuous process features as the bottleneck of such algorithms.According to the time-domain characteristics of data acquisition requirements,a high-frequency data acquisition system was selected.The data was further pre-processed,most of the valid data sets were retained,and the stages were divided according to the data characteristics of the human pose.Hidden Markov theory is studied,and different order Markov models for attitude recognition are constructed.A Gaussian mixture model is used to fit the state of each stage that the human pose cannot directly observe,and a human pose recognition model is constructed based on Hidden Markov.The performance of the GMM-HMM model and other algorithms on this dataset is compared and compared..The applicability of hidden semi-Markov models in human pose recognition was studied.In view of the fact that the traditional hidden Markov model cannot pay attention to the time characteristics of the attitude,a GMM-HSMM model is constructed.The neural network was used instead of the Gaussian mixture model,and the DNN model structure was explored.Finally,the best model structure,activation function type,initialization method,and gradient descent method were determined.Finally,the performance of the previous chapters and other algorithms are compared.(3)This paper designs a data acquisition system and a human pose recognition algorithm based on the characteristics and behavioral characteristics of the human body.This paper starts from the two aspects of the instantaneous characteristics and the time domain characteristics of the sensor data corresponding to the human pose,and reasonably configures the hardware system to provide reliable support for algorithm modeling.In this paper,four algorithms for human body recognition based on multi-sensor signals are designed,and the performance of the model is evaluated in terms of recall rate,accuracy rate,and accuracy rate.Eight gestures can be identified respectively,and the recall rate of the optimal model is 98.6%.