Research On Multi-sensor Human Activity Recognition Based On Residual Network

With the development of the Internet of Things,the concept of human-computer interaction has been further expanded.Human activity information needs to be collected in a timely manner and effectively processed and recognized to further assist and support human behavior activities.Therefore,how to be more effective in the current context Multi-sensor human activity recognition is a very important research direction.Researchers have conducted a number of studies on simple movement patterns and complex activities over long periods of time,but there are still few studies on interactive activities based on multi-sensors.Multi-sensor interactive activities have the characteristics of transientity,continuity,and interactivity.Models suitable for simple motion patterns are often not sufficient to characterize their rich activity characteristics;while complex activity models with relatively long periods of time,the details of interactive activities The granularity is higher,the time span is smaller,and it has a certain degree of recognition difficulty.For multi-sensor interactive activity recognition,the existing algorithms still need to be improved and performance needs to be improved.Therefore,this paper studies multi-sensor interactive activities.The main work of the paper is as follows :1.The multi-sensor human activity recognition technology is analyzed and summarized.The thesis expounds and summarizes the development background of human activity recognition,the research status at home and abroad,and introduces various related theories of human activity recognition.2.Multi-sensor interactive human activity recognition involves the fusion of people and objects,and people and the environment.Aiming at the characteristics of spatiotemporal interactivity,ephemerality and continuity unique to multi-interactive activities,this paper proposes a human activity recognition algorithm based on window preprocessing and grouped residuals combined with spatial learning.On the one hand,the experimental analysis and comparison of the effect of sliding window processing on human activity recognition during data preprocessing,includingdifferent sliding window sizes and coverage,etc.;on the other hand,based on multi-sensor interactive activity recognition Window pre-processing conclusion,further use grouping residual joint spatial learning algorithm for activity recognition and classification,and design multiple sets of comparative experiments to optimize the network model,loss function,classifier,etc.respectively.Finally,a comparative experiment was conducted on the Opportunity activity data set,and the algorithm performance exceeded the related state-of-the-art activity recognition algorithm,which further verified the multi-sensor human activity recognition based on window preprocessing and grouped residual joint spatial learning.The effectiveness of the algorithm.3.Combined with window preprocessing,grouped residuals combined with spatial learning algorithm,the paper further develops a multi-sensor based human activity recognition system.The recognition system can be divided into two subsystems: classification system and background management system.Among them,the classification system is mainly used for ordinary users to use classification algorithms for activity data,and the background management system is mainly used for visual management of background data by non-development management personnel.