Human Motion Pattern Recognition Research Based On Random Forest Algorithm

With the development of sensor technology and the continuous improvement of pattern recognition algorithms,sensor-based human motion pattern recognition technology has become a field of concern and develops rapidly.Now it has been applied to human-computer interaction,health monitoring,sports competition,Military and many other fields.The basic principle of human motion pattern recognition is to determine the type of motion by processing and analyzing the human motion information acquired by the sensor.It's main steps includes data preprocessing,feature extraction,feature selection,classifier training and classification algorithms.In order to limit the computational cost of the system and to obtain enough data to ensure the accuracy of the system classification results,this paper selects the data of sensors that located on the wrist and chest to participate in the recognition of human movement patterns.Due to the complex human movement and the multiple changing of environment,there are still many problems exist in human motion pattern recognition.For example,more meticulous classification of different sports patterns,more effective feature extraction and more efficient classification algorithms are still needed.Focusing on these difficulties,this paper starts the researching on human pattern recognition system,The main work includes:The UCI database PAMAP2 data set is selected as raw data to paticipate in experiments,that data set is characterized by comprehensiveness and rich data points.In the data preprocessing part,the Kalman algorithm parameters are optimized and a Kalman filter suitable for human motion pattern recognition is realized,then use that filter to denoise and filter the original data.A multi-feature fusion algorithm based on discrete wavelet transform and autocorrelation function is proposed and used it to process and analyze the data.The algorithm uses the discrete wavelet transform to do multiresolution wavelet decomposition of the data signal and extracts the eigenvalues at each resolution.At the same time,the autocorrelation function is used to transform the data,and the features that can reflect the cyclical characteristics of the motion time domain are extracted.Combining the two parts of eigenvalues to generate eigenvectors,and then using local linear embedding method to reduce the dimension of the vectors to complete the fusion of features.An algorithm names Artificial Bee Colony Optimize Random Forest is proposed,and used as classification in human motion pattern recognition.Firstly,choose CART algorithm to apply to random forest decision tree;then based on Artificial Bee Colony Optimize Random Forest Algorithm to optimize parameters such as maximum depth and maximum number of random forest algorithm.Experiments show that human motion pattern recognition based on that algorithm can get higher classification accuracy.Based on the human motion pattern recognition algorithm above,a human motion pattern recognition system based on B/S structure is designed and implemented.The system can provide valuable exercise pattern information for those engaged in human-computer interaction,sports,health care and other related work,it allows users to share the data of exercise patterns and provide technical support for the development in the field of pattern recognition.