Human Activity Recognition Method Based On Triaxial Accelerometer

With the development of microcomputer technology,smart mobile devices and various wearable sensors are widely used.As a revolutionary technology,wearable technology has played a pivotal role in health care,sports and health,smart cities,and military operations.Based on the rapid development of wearable technology,user context awareness,as an important field in pervasive computing,has received more and more attention.In recent years,research on the use of wearable devices and smart mobile devices to recognize human activities has attracted more and more attention.As an extension of human activity recognition,crowd behavior recognition plays an important role in transportation and traffic planning,urban public area planning,service recommendation,crime prevention,and social event monitoring.However,due to limitations in battery capacity and computing resources,context-aware applications such as activity recognition have not been widely accepted by the general public.In order to meet the energy-efficient requirement of mobile devices application,this paper first proposes a lightweight human activity recognition algorithm based on activity pre-classification and historical activity analysis.The algorithm uses the triaxial acceleration data to extract the time domain features,and reduces the execution frequency of the classification algorithm with high complexity.In addition,based on changes in user activity,the lightweight algorithm adjusts the sensor sampling frequency in real time to reduce power consumption and decrease system response time.On the other hand,the traditional group behavior recognition methods mostly use surveillance cameras to collect video data to recognize group activities,which would have the problem of lower coverage and infringement of user privacy.Therefore,this paper presents a method for recognizing group behavior using the triaxial accelerometer.This method only needs to collect user 's motion data,and obtain the behavioral correlation matrix of the population through the feature level fusion method.According to the different scale of the population,the multi-dimensional scaling method or spectral clustering algorithm are used to divide the group behavior pattern.Experiments show that the lightweight human activity recognition method proposed in this paper achieves an accuracy of 92.5% for the recognition of 8 human daily activities.Compared with the traditional SVM algorithm,under the same conditions,the energy consumption of the lightweight algorithm is reduced by 49% and the delay is reduced by 55%.Further,we conducted group behavior recognition experiments among groups of 10 users.The results show that the group behavior recognition method proposed in this paper can effectively divide five different behavior patterns in the group.For walking,running and other dynamic activities,the accuracy of group behavior recognition reached 90%.