Research On Key Technology Of Human Activity Sensing Based On Wireless Communication Signal

As a key technology for intelligent interaction,human activity sensing plays an important role in improving the intelligence and convenience of people’s life.Compared with the traditional sensing methods based on wearable sensors,computer vision,and radar,the emerging human activity perception technology based on Wi-Fi wireless communication signals has advantages,including obviating the need for carrying devices,making full use of the deployed communication infrastructure,not limited by light conditions,providing better protection of users’ privacy,etc.With the increasing research work,wireless intelligent sensing technology based on Wi-Fi signals has achieved fruitful achievements,but there are still many difficulties and challenges.For instance,low accuracy for location-unfixed sensing,weak location-adaptive ability,strong dependence on the number of samples,limited flexibility of activity category updating,etc.To address the above issues,this paper carries out in-depth research on the key technologies of human activity perception based on Wi-Fi signals,and takes full advantage of the powerful feature extraction ability and identification capabilities of deep learning methods,aiming to improve the accuracy,generalization,and scalability of wireless intelligent perception methods under small sample conditions.Specifically,the following innovative research results have been accomplished:(1)Location-unfixed human activity recognition methodHuman activity recognition technology involving fixed location is difficult to meet the needs of practical application scenarios.Locationunfixed sensing is an inevitable requirement of wireless intelligent sensing.To solve the problem that the accuracy of location-unfixed perception is unsatisfactory,a sensing method based on amplitude and phase enhanced deep complex network is proposed.High-accuracy location-unfixed activity recognition is realized by mining abundant activity information.The proposed method is validated with 5 human activities from 24 sampling locations in a real office environment and achieves 96.85%average recognition accuracy.On this basis,considering the unbalanced number of training samples in different positions,we propose a sensing method based on a complex convolution transfer network to apply the empirical knowledge learned from part locations with sufficient activity samples to other locations with insufficient samples.Experiment results demonstrate that the proposed method can achieve an average recognition accuracy of 94.02%in the case of unbalanced samples.(2)Location-adaptive human activity recognition methodWeak location-adaptive ability is a key problem to be solved in the field of wireless intelligent sensing.Especially,when the available data samples are limited,it is impossible to learn the specific feature representation of each human activity at different positions through model training.It is necessary to learn a location-adaptive activity discrimination method to improve position generalization performance.This paper proposes a location-adaptive human activity perception method that pays more attention to the metric relationship between activity categories and weakens the location information.A dual activity feature representation method with sample serialization based on a convolution neural network and long short-term memory network is proposed to mine the discriminative features of different activities and the common features of the same activities at different locations.Furthermore,the recognition method improved by metric-based meta-learning aims to increase the interclass distance and reduce the inner-class distance.By measuring the similarity of activity categories,the samples from the same activity category at different locations can be gathered adaptively in the feature space,which promotes the discrimination accuracy of activities at different locations and realizes the location-adaptive activity perception.Experiment results show that the proposed method can obtain 91.11%recognition accuracy of four human activities when only one sample is provided for each testing location.In addition,it can still obtain a good sensing effect when the data sampling rate is low and the number of transceiver antennas is limited.(3)Location-independent human activity recognition methodThe dependence of the wireless intelligent sensing method on data sample size greatly limits its popularization in practical applications.Particularly,for perception scenes with high requirements for location generalization,only data samples involving partial locations can be provided,which will seriously limit the accuracy of the perception method.This paper further explores the location-independent human activity recognition method with very few training samples.We propose a feature extraction method based on the attention mechanism,which focuses on the location-independent feature in the three dimensions involving the channel,time,and subcarrier.In addition,a human activity perception method based on few-shot learning is proposed.By designing similar training and testing tasks,the model is guided to learn location-independent features.Experiments show that the proposed method can obtain 91.98%recognition accuracy of four human activities when only five samples are provided for each activity,and it has good superiority and robustness.The proposed method effectively alleviates the dependence of the human activity perception method on the number of training samples.(4)Class-incremental human activity recognition methodExisting wireless intelligent sensing methods mainly focus on recognizing the predefined fixed activity categories or uniformly identifying the undefined new categories as class "other",rather than identifying new categories one by one.Especially,considering the limitations of storage space and computing resources of perceptual terminal devices,it is impossible to provide adequate samples of the old activity category.Requiring users to provide plenty of samples of new activity categories can be very disruptive to the user experience.To solve these issues,a method of small sample class-incremental human activity recognition is proposed.In order to extract more abundant features from limited samples,a feature extraction method based on amplitude and phase enhanced convolutional neural network is proposed.And a double loss function design scheme based on a complete subcarrier set and compressed subcarrier set is designed.A phased model training and updating strategy are provided to overcome the knowledge forgetting of the old categories.In order to verify the performance of the method,10 daily behavior activity data were collected in three environments,respectively.The proposed method can obtain 85%recognition accuracy for 5 old categories and 5 new categories,with each new category providing only 5 samples.The proposed method realizes the class-incremental human activity perception under the condition that both old and new activity samples are limited,and provides a new idea for improving the category scalability of the human activity perception method.