Research On Human Activity Recognition Algorithm And Its Application Research In Smart Prison

Human activity recognition(HAR)is an active field in ubiquitous computing and body area network(BAN),which has been widely applied in medical care,sport and smart home.With the development of computer and microelectronic technology,various types of sensor devices are beginning to appear in all aspects of people’s lives.The sensor devices are small in size,fast in calculation speed and portable,which makes the research of HAR based on sensors get more and more attention.At the same time,machine learning,deep learning and other artificial intelligence algorithms show good performance in HAR,which brings new opportunities to the development of HARAfter decades of research by researchers,HAR technology has made great progress,but there are still many problems and challenges.For example,in terms of feature extraction and classification,traditional algorithms still have problems such as incomplete feature extraction,low accuracy when identifying similar and complex activities,and the lack of a general recognition framework.Regarding the application of HAR,the attention of abnormal activities of special social groups is insufficient and related datasets and researches are still lacking.In view of the shortcomings of existing methods and researches,the research contents of this thesis are as follows(1)A sensor-based activity recognition system framework is proposed,which can be applied to sensor-based HAR tasks in general.At the same time,a HAR network model based on spatiotemporal multi-feature extraction(SMFE)is proposed.The model can use the temporal and spatial features to distinguish similar activities and improve the accuracy of complex activity recognition effectively.(2)The attention mechanism is introduced into the HAR system.On the basis of squeeze-and-excitation(SE)block,a spatial and channel based SE(SCbSE)block is proposed,which can recalibrate features in two dimensions:space and channel.The SCbSE block can be directly added to any Convolutional Neural Networks(CNN)layer to further improve the accuracy of activity recognition(3)Regarding the actual needs for recognizing aggressive activities,this thesis simulates the prison environment and collects an aggressive activity recognition dataset(AARD),which includes three common aggressive activities and five daily activities,to make up for the lack of abnormal activity recognition datasets.At the same time,an aggressive activity detection method based on threshold is proposed according to the characteristics of the aggressive activitiesThis thesis experiments on the WISDM,OPPORTUNITY and AARD datasets.The results prove that the proposed SMFE model can effectively improve the accuracy rate and better distinguish similar activities;the proposed SCbSE block can be effectively embedded into the existing network and further improve the recognition accuracy;the proposed aggressive activity detection method based on threshold can simplify the model and improve the recognition speed while ensuring the recognition rate.Through the above research,the performance of the HAR system to recognize similar and complex activities is improved,and the application of HAR in smart prison scenarios becomes more practical.At the same time,the system framework and related improved algorithms mentioned in this thesis can be extended to other application scenarios,which has great application space and research value.