Human Activity Recognition Based On Deep Learining And Wearable Sensors

With the continuous development of microelectronics and sensor technology and the continuous development of deep learning theory,human activity recognition based on wearable devices has become a new research direction.This research can better reflect the essence of human motion.Compared with human activity recognition based on computer vision,human activity recognition based on wearable devices is not limited by specific scenes and time and that consumes less energy and has lower cost,so it is more suitable for promotion.Although the development of human activity recognition based on wearable devices has made great progress in recent years,more problems are still urgently needed to be solved,such.as how to extract the features with strong characterization ability,how to design a high-precision sensor data acquisition system and how to design an end-to-end human activity recognition system.This thesis is different from the traditional features of artificial extraction,using the deep learning technology to carry out the following work:The first is to use convolution feature extraction methods to build an end-to-end human activity recognition system.This system is based on different convolution feature extraction methods(one-dimensional convolution feature extraction method,two-dimensional convolution feature extraction method,and one-dimensional convolution combined with the recurrent feature extraction method).We construct three different human activity recognition models.And then we verify and analyze the experimental data on our pre-processed samples to obtain the recognition effect of each method and compare their respective advantages and disadvantages.The second is to use recurrent neural networks to build an end-to-end human activity recognition system.Through a simple Recurrent Neural Network(RNN),a Long Short-Term Memory(LSTM),a Bidirectional Long Short-Term Memory(BLSTM)and Gated Recurrent Unit(GRU),we construct four different human activity recognition models.And then we also perform experimental verification and analysis on our pre-processed sample data,finally we obtain the recognition effect of each method and compare their respective advantages and disadvantages.The third is based on Android system to design a sensor activity motion acquisition software.We can use this software to collect activities which we want to collect,such as cycling,making calls,eating breakfast and so on.At the same time,we can specify the time and frequency of collection.The collected activity samples can be single activity data with strong labels.It also can be multiple activity data with weak tags.