Font Size: a A A

A Research And Implementation Of Human Activity Recognition System Based On MapReduce And Intelligent Mobile Device

Posted on:2018-11-11Degree:MasterType:Thesis
Country:ChinaCandidate:R C LiFull Text:PDF
GTID:2348330515951794Subject:Software engineering
Abstract/Summary:PDF Full Text Request
In recent years,human activity recognition technology based on sensor data has been widely used in many fields such as,motion monitoring,virtual reality and so on.However,due to the sharp increase of intelligent mobile devices,the amount of data it needs to process is also growing rapidly.It has been unable to meet the current needs if we still use the traditional standalone mode to process data,build classification model and recognise the unknown behavior.On the basis of existed research,this thesis analyzes the current scheme of human activity recognition and classification algorithm,and propose a new recognition scheme based on MapReduce framework and intelligent mobile device.Meanwhile,in order to realize multi-user realtime activity recognition,including sitting,standing,walking,running,upstairs and downstairs,this thesis chooses Flume,Kafka and Storm framework to build the real-time activity recognition system.The core of this thesis mainly focuse on two aspects: offline data modeling and multi-user realtime activity recognition.Offline data modeling is mainly responsible for the parallelization of decision tree algorithm based on Map Reduce framework,and uses the training set to build activity classification model.Firstly,the system acquires user activity data through the built-in triaxial acceleration sensor of the smartphone and uses sliding window to divide the continuous data.Meanwhile,in order to obtain the real human activity data,this system uses the low-pass filtering algorithm to isolate the force of gravity.Then it uses a moving average filter to remove unwanted noisy component from sensor data.After the preprocessing operation,according to the characteristics of the six kinds of behavior data,this thesis uses a 22-dimension eigenvector to describe the individual behavior fragment.Finally,this thesis improves the traditional C4.5 decision tree algorithm based on the MapReduce framework.By using the map and reduce function,it computes each attribute information gain ratio in a parallel fashion and choose the best splitting node to build decision tree model.Based on the experimental results,the accuracy of the recognition rate of the above six kinds of movements can be achieve 86.63%.The algorithm performance has been greatly improved and suitable for mass data processing.Multi-user realtime recognition is mainly responsible for data transmission and collection,and the construction of topology task.At first the client transmits data to be recognised through the http protocol.And the back-end uses the Flume cluster and Kafka cluster to collect data.Then it uses the kafka cluster to provide real-time data stream.Finally,the Topology task is constructed.The spout node is used to read data from message queue.And the bolt node is used to classify the activity eigenvector based on the constructed decision tree model.Then it aggregates recognition results and stores the data in mysql database.According to the experimental results,The real-time computing system has high throughput and performance.And it can effectively process a large number of users' realtime activity data.
Keywords/Search Tags:human activity recognition, decision tree, MapReduce, realtime computing, intelligent mobile device
PDF Full Text Request
Related items