Research On Human Activity Recognition Based On Deep Learning

In recent years,with the rapid development of smart devices,the range of human activity recognition applications has been continuously expanding.However,due to the problems of low efficiency and low accuracy of traditional manual feature extraction,recognition of human activity based on deep learning becomes a new direction for research.Compared with traditional methods,accuracy and efficiency have greatly improved.However,there are still two problems in the research.On the one hand,there are fewer open datasets,which results in insufficient data samples in the study,and insufficient dataset concentration.On the other hand,the accuracy rate of current algorithms in the field of activity recognition has reached saturation state,but the accuracy still has room for improvement.In order to solve the problem of insufficient data samples,this thesis expands and reconstructs the seven published datasets to generate the target dataset SADataset.SADataset contains 14 activity categories,more than 53460×200 data points,and covers most of the activities of daily life.It is very suitable for scientific research in this thesis.In this thesis,the activity recognition problem is regarded as the dual problem of image recognition problem.The activity is the lable of the image.The sensor dataset in the time domain is transformed into the image space by using the Gramian Angular Fields data processing algorithm,the Markov Transition Field data processing algorithm,and the Spectrogram data processing algorithm to obtain the activity recognition image dataset SAImageNet.This thesis analyzes the SAImageNet dataset and finds that there are certain similarities in the images generated by the same kind of activity,and there are certain differences in images generated by different activities.SAImageNet datasets are used to do multi-classification tasks,and convolutional neural networks is improved in three aspects.Firstly,in order to quickly screen out high-value information from the input data,attention mechanism is introduced in the third-dimensional feature channel,we add a weight to each feature channel to amplify useful channel information and suppressing invalid channel information;Secondly,a gate control mechanism is introduced to solve the problem of partial gradient disappearance.Finally,we introduce the residual block to solve the problem of the disappearance of the gradient.Combining the above three aspects of improvement,this thesis proposes an AFRRNet model for adaptive feature recalibration of residual network,training the model on the target dataset SADataset,adapting the model to data,and possessing generalization capability.Finally,in order to verify the validity of the AFRRNet model,the CNN model,RNN model and AFRRNet model were trained on the generated image dataset SAImageNet.The experimental results show that the AFRRNet model has increased the left right of one percentage point compared to the other model recognition accuracy,and achieved the expected experimental results.