Design And Implementation Of An Activity Recognition System Based On CNN

With the progress of microelectronic systems and sensor technologies,sensor-based behavior recognition research has attracted the attention of researchers.Ubisense positioning system based on UWB technology has been applied to the field of behavior recognition because of its high positioning accuracy and short positioning time.Therefore,to build the Ubisense based behavior identification system is theoretically significant and has a broad application prospect.By wearing multiple Ubisense tags(UWB-RFID)at key parts of the user's body,real-time spatial location information for body parts under different behavioral conditions can be obtained.On this basis,this paper adopts the machine learning method to realize the identification of daily behavior for running,jumping,walking,standing,sitting and lying,it falls into following aspects:Firstly,the effects of different models such as CNN(Convolutional Neural Network)algorithm and BP(Back Propagation)neural network algorithm on behavior recognition are studied.A large number of experimental results show that CNN model is superior to BP neural network model,and the accuracy of behavior recognition is 87%.Secondly,by selecting the number and location of different tags to wear,the impact of the wearing scheme on behavior recognition is analyzed.The experimental results show that,in terms of the number of tags,we can appropriately reduce the number of tags without affecting the overall behavior of identification;in terms of the location of tags,wearing the wrist-waist-ankle recognition the best.Thirdly,based on the research results of the CNN model,an online behavior recognition demonstration system is realized by J2EE and the SSH framework.In summary,this paper studies the behavior recognition system based on Ubisense positioning system,and train the CNN model for identifying six kinds of common behaviors through a large number of experiments,and develop a Web system for demonstration.