Design And Implementation Of Pedestrian Transportation Action Recognition Based On Mobile Phone Sensors

The goal of Human Activity Recognization(HAR)is to identify common human activities in real life.The research began in the 1980 s and was mainly used in medical and health care.Pedestrian transportation action recognition is a kind of HAR in which identify stilling,walking,running,cycling,riding cars,buses,subways and trains is needed.With the popularity of smart phones,research on the recognition of pedestrian traffic behavior based on various sensors built into mobile phones has attracted more and more attention.At present,the recognition accuracy of walking and running reaches more than 95%,but the accuracy of recognition in various vehicles and stationary state is low.In this thesis,a pedestrian transportation action recognition method based on two-level random forest classifier is proposed for the above problems,then the performance of common classifiers is analyzed.Finally,the accuracy of identifying various vehicles and stationary states is improved using a small computational cost.By analyzing the influence of random forest parameters on classification accuracy,Bayesian optimization algorithm is chosen to optimize parameters.In the feature extraction stage,the defects of existing features in the identification of vehicles is analyzed,and a method for extracting energy features in the frequency domain interval is also proposed.In addition,according to the conversion relationship between states in the real scene,the hidden Markov model whose initial transfer matrix is setted also added to establishes a state time series to predict the current state,corrects the final classification result,and improves the classification accuracy in the real scene.The results show that the classification accuracy of the stationary and the vehicle states can be increased to more than 95%,and it is robust to different pedestrians and different mobile phone placement positions.