Research On Human Activity Recognition Algorithms For Internet Of Things Devices

With the development and popularization of Internet of Things technology,Internet of Things devices are playing an increasingly important role in people’s life and work.Among the many Internet of Things devices,smart wearable devices and smart home devices have become an essential part of people’s lives.These devices can collect vast amounts of sensor data,including data on human movements.Human activity recognition technology can process and analyze these data,so as to achieve human-computer interaction,health monitoring,behavior analysis and other application scenarios.Therefore,human activity recognition technology has important application value in Internet of Things devices.Traditional human activity recognition methods are mainly based on hand-designed feature extraction,but there are some problems such as inadequate feature expression and weak generalization ability of classifier.The rise of deep learning technology has brought a new breakthrough for human activity recognition.Neural networks can automatically learn more effective feature expression and adapt to more complex classification problems,so it has become a mainstream method in the field of human activity recognition.For most existing neural networks,a large amount of label data is needed to train a model with good classification effect,and the data is expected to be balanced in categories.In a real-world scenario,however,such data is not readily available.This research puts forward some solutions to such problems,and the specific work is as follows:(1)An integrated model based on Autoencoder is proposed to realize human activity recognition.Aiming at the problem that there are some small sample data in human activity datasets,which makes model training difficult,an Autoencoder-based human activity recognition technology is proposed.This technique relies on the reconstruction ability of the relevant data of the Autoencoder and takes the reconstruction effect as the classification basis,so that the model training effect is not affected by the number of small sample data.The experimental results show that this method can improve the poor recognition ability of neural networks for small sample data.(2)The use of adversarial autoencoder and its improved structure for human activity recognition is proposed.This research point mainly aims at the problem of insufficient labeled data in human action datasets,and proposes a structure combining autoencoder and generated adversarial network.The adversarial autoencoder is used for semi-supervised learning to complete the task of human activity recognition when the proportion of labeled data is relatively low.Combined with the characteristics of sensor data,the algorithm is improved.Experiment results show that the adversarial autoencoder and the improved convolutional adversarial autoencoder achieve better classification results when the proportion of label data is very low.(3)Designed and implemented a system to manage human activity recognition algorithms.The management system provides an algorithm management platform which is easy to use and easy to read.The models,datasets and training parameters are centrally managed,and the experimental steps and results are visualized.