Human Activity Recognition Algorithm Based On Inertia Measurement Unit Sensor And Deep Learning

Human Activity Recognition(HAR)can be defined as determining a person’s various postures and daily activities through a series of observations and surrounding environments.The HAR technology based on the Inertia Measurement Unit(IMU)sensor has attracted the extensive attention from numerous researchers because of its advantages such of low cost,small size,and fortissimo resistance to environmental interference.In recent years,many scholars have applied deep learning technology in HAR.However,existing deep learning-based HAR models often have high complexity,large computing power requirements,and poor generalization and robustness.To address these issues,this paper focuses on HAR methods based on IMU sensors embedded in smartphones,and integrates various technologies such as deep learning,metric learning,and ensemble learning.The aim is to further reduce the complexity and computational cost of the HAR model,while improving their recognition performance and out-of-distribution generalization performance.First,a lightweight Human Activity Recognition(HAR)algorithm based on Residual MultiLayer Perceptron(Res-MLP)and Gaussian Error Linear Unit(GELU)activation function is proposed to address the issue of high complexity in commonly used HAR algorithms based on IMU sensors.Experimental results on the publicly available UCI HAR dataset demonstrate that the proposed method not only has better performance,but also has lower complexity.Furthermore,in order to verify the generalization ability of the algorithm,this paper re-divided the dataset and created four out-of-distribution(OOD)datasets for testing.The results demonstrate that the proposed algorithm has strong generalization ability,indicating its potential for practical applications.Then,in order to improve the out-of-distribution generalization of the Res-MLP algorithm,this paper proposes a HAR method named RMDML-HAR,which combines Res-MLP and deep metric learning feature embedding technology to extract generalized and discriminative features,thus improving model recognition performance and generalization ability.This method utilizes the ResMLP network as the backbone network for feature extraction,and combines two contrasting losses and softmax loss to propose a hybrid metric loss function aimed at discriminative feature embedding,ensuring that the intra-class distances in the feature space are compact while the inter-class distances are separable.Experimental results on the four OOD datasets demonstrate that the RMDML-HAR method effectively improves the Res-MLP model’s OOD generalization performance and classification accuracy.Finally,to further improve the recognition and generalization performance of the model,this paper introduces an ensemble learning approach using Stacking on the basis of the RMDML-HAR method,proposing the RMDMEL-HAR algorithm for human activity recognition based on Res-MLP and deep metric learning ensemble learning.The base learner uses the RMDML method,and the meta-learner uses the support vector machine algorithm in machine learning.This not only ensures the improvement of recognition and generalization performance but also maintains the low complexity of the model.Experiments are carried on four OOD datasets,to compare the performance with the single classifier method,verifying the improved recognition and generalization performance of the proposed RMDMEL-HAR method.