Research And Implementation Of Human Activity Recognition Transfer Technology For Smart Device

With the continuous development of ubiquitous computing and mobile intelligence,activity recognition has been widely applied in people’s daily lives.The increasingly complex behavior demands and inconsistent data distributions in different environmental semantics pose challenges for building efficient activity recognition systems.Therefore,this dissertation focuses on human activity recognition transfer technology for smart device,with a focus on breaking through the following challenges:insufficient learning capability for complex behavioral semantic feature representation,differences in data feature distribution across locations,and difficulty in generalizing unknown user models.This dissertation addresses these challenges by carrying out the following work:(1)To address the issue of insufficient learning capability for complex behavioral semantic feature representation,this dissertation proposes a activity recognition method based on multi-scale spatial semantic mining.The method first extracts and expresses semantic spatial information using multi-scale time windows,expanded the representational capacity of behavioral semantic information.Next,the key semantic information of elevator,escalator,and stair behaviors is analyzed.A robust feature extraction algorithm is designed using the relative variability of spatiotemporal features.Furthermore,an ensemble learning algorithm is designed to enhance the model’s generalization performance,while ensuring model stability,effectively improving the performance of complex behavior recognition;(2)To address the problem of differences in data feature distribution across locations,this dissertation proposes an activity recognition transfer method based on subspace adaptive alignment.The method first designs a sensor-based multi-scale feature extraction model,SMSNet,which learns multi-scale features from low to high scales in different feature spaces,effectively integrating semantic representations in different scale spaces,and obtaining more efficient and robust higher-order features.For the feature distribution alignment problem,a domain alignment module is designed to adaptively align marginal distribution features,and a subspace alignment module is designed to adaptively align conditional distribution features;(3)To address the difficulty of generalizing unknown user models,this dissertation proposes an activity recognition transfer method based on style semantic contrastive generalization.The method first proposes a multi-domain feature stylization module to generate diversified style samples,greatly expanding the style representation space of the source domain.After incorporating diversified style samples into training,in order to maintain the continuity of semantic information,semantic consistency regularization loss is used to maintain semantic consistency between generated and original samples.In addition,since feature generation inevitably mixes style noise in the semantic representation space,this method also designs a style semantic contrastive loss to maximize the spatial distribution of semantic diversified features,making the model insensitive to cross-domain style variations and more focused on crossdomain semantic invariance information.Experimental results show that the above three methods are superior to existing activity recognition and transfer learning methods,achieving higher accuracy and better generalization performance.Furthermore,this dissertation designs and implements an activity recognition transfer system for smart device to achieve efficient data collection and precise transfer of activity recognition models.