Research On Activity Recognition Analysis Based On Intelligent Perception

The growth of wearable devices has brought out the increase of the number of IoT(Internet of things)devices and data year by year,which on one hand pro-motes the development of digital economy,but on the other hand,increases the pressure of cloud computing.In the 14th Five-Year Plan,our government em-phasizes promoting the development of new infrastructure and accelerating the creation of new advantages of digital economy.With the development of new infrastructure,mobile edge computing eases the pressure of cloud computing,and promotes the development of digital economy by using emerging technolo-gies such as big data and IoT.As the representative of mobile edge computing,wearable devices equipped with various sensors are infiltrating people's lives.Among the computing systems of these equipments,there is a basic and impor-tant technology-human activity recognition technology.Human activity recognition technology aims to analyze and recognize the current behavior state of the human body through perceiving the movement and physiological information of the human body by various sensor devices.Human activity recognition,as an important branch of pattern recognition and perva-sive computing,has been widely used.According to the information source,human activity recognition includes vision-based human activity recognition and sensor-based human activity recognition.With the development of micro-electronics technology,small volume and low power sensor equipments have begun to penetrate into people's lives.At the same time,with the help of the powerful analysis ability of artificial intelligence,sensor equipment has acceler-ated the development of intelligent perception.Therefore,sensor-based human activity recognition has more potential.However,the deployment of perceptual algorithm in wearable devices requires higher recognition performance,prac-ticability,lightweight deployment and robustness.Therefore,based on intelli-gent sensing technology,this paper studies sensor-based human activity recog-nition,and provides theoretical support for the establishment of high-precision,practical,lightweight and robust human activity recognition system.Briefly as follows.First,aiming at the difficulty of feature extraction in the task of activ-ity recognition,we propose a high-precision perception model.In this model,we design the dense connection module and multi-layer feature aggregation module to ensure the maximum information flow in the network,extract the high-dimensional features of time dimension and sensor dimension,and fuse the multi-layer feature information according to the feature importance.The experimental results show that compared with other human activity recogni-tion models,the high-precision perception model does better in classification performance.In addition,ablation and visualization experiments verify the ef-fectiveness of the proposed two modules.Second,aiming at the degradation of classification performance in open-set recognition,we propose an open-set recognition framework.The recog-nition framework introduces a margin mechanism into the softmax function,which makes neural network learning have more differentiated high-dimensional features,thus increasing the difference of the features between classes.The simulation results show that the neural network model trained by the improved soflmax function has better classification effect than which trained by the tradi-tional softmax function,and the open-set recognition framework can effectively solve the problem of unknown category in the human activity recognition.Third,aiming at the lightweight requirements of wearable device system deployment,we propose a lightweight perception model search scheme.The search scheme is based on the neural architecture search theory,using NSGA-II algorithm to find neural network model under multiple search targets,and provides theoretical reference for the design of automation networks of human activity recognition.The experimental results show that the network searched has lower rate in classification error and computational complexity than other human activity recognition networks.Fourth,aiming at the poor robustness of human activity recognition based on single mode,we propose an enhanced-perceptual algorithm based on multi-modal data.The enhanced-perceptual algorithm uses inertial sensor data and skeleton data.The inertial sensor data is processed by dilated convolution neural network and skeleton data is constructed into a Gaussian heatmap con-taining time series information,which is sent to convolution network to ex-tract features.Finally,the enhanced-perceptual algorithm uses different fusion strategies to complete the classification.Experimental results show that the enhanced-perceptual algorithm can significantly improve the classification per-formance compared with the single modal algorithm.