Research On Daily Physical Activity Recognition Based On Multi-classifier

Scientific and moderate physical activity plays a vital role in maintaining human health and preventing various chronic diseases.However,the fast-paced and high-intensity working environment in modern society makes people lack of exercise,which directly impacts the health of people.Therefore,it is of great research significance to track and quantify the daily physical activities.Human activity recognition based on wearable device can eliminate the limitation of the occasion and shows broad application prospects in the fields of medical rehabilitation and fitness management.In this study,while optimizing the existing ensemble model based on multiclassifiers,a cascade-classifier based on multi-label learning is proposed to realize effective recognition of daily human activities.The main contents of the study are as follows:(1)The processing methods of sensor signal in different nodes of human body are analyzed.Based on data processing operations such as feature extraction and selection from different dimensions,the obtained activity instances can effectively represent different types of activities.In addition,the conventional cross-validation method is improved as well to reduce over-fitting of results and ensure the reliability of the model.(2)The differential ensemble model is proposed and constructed from three aspects: differentiation training-set combination,differentiatial feature selection and differentiatial voting weights.Through the level of data fusion,the data from each sensor node is combined to form a new training set.Selecting different features by randomly and a weighted voting method based on activity type is proposed.The results of the experiment show that the classification performance of the differential ensemble model is significantly better than that of the conventional ensemble model.(3)A cascade model based on multi-label learning has been proposed as well for the human activity recognition.Specifically,based on sensors of different moduals,activity intensity labels are introduced to represent daily human activities,and cascade models are constructed to refine the attention of the base classifier to different modal attributes step by step.Also tested on human activity data sets,the multi-label cascade model achieves higher activity recognition accuracy with less training time consumption and a smaller model structure.(4)A human activity recognition platform has been developed,which realizes the functions of data reception and storage,feature extraction,data format conversion,etc.A dynamic interactive web page is designed to realize real-time display of sensor data and real-time feedback of predicted activities.