Research On Key Issues Of Sensor-based Activity Recognition

With the rapid development and gradual maturity of low-power micro-sensors technology,the application system based on sensors,of course,is common enough in day-to-day life,providing substantial facilitation for our working and living.Meanwhile,the method based on sensor,which is different from computer vision-based activity recognition,much better reflects the essential features of human activity,without being confined to specific scenarios and time.Besides that,the method protects users’ privacy,makes data freely available and acquire a large amount of information.As it turns out,activity recognition based on sensors that owing small volume,high sensitivity along with simple equipment,has become a new research hotspot and obtained scholars’ both at home and abroad high attention.Although activity recognition based on sensors has made a great progress in recent years,many problems still remain unsolved.The thesis focuses on the following issues:data segmentation,optimal feature selection and data fusion.Firstly,aiming at the recognition problem of various patterns,such as the periodicity of the same kind of activity,the alternation of different kinds of activities and the transitional one when the activity changes,the thesis proposes a data segmentation method,which implements the accurate segmentation of data of basic activity and transitional activity in multi-modal activities,based on Mahalanobis distance.At last,the validity of the proposed method is verified by comparing the sliding window with a fixed size and no overlap,the window with a fixed size of 50%overlap and the data segmentation method proposed in this thesis.And then,in view of the issue that a majority of traditional feature selection methods take insufficient account of the correlation of features,the thesis puts forward the feature selection method related to feature relevancy based on the Binary Particle Swarm Optimization(BPSO)considering the recognition rate as a criterion.To increase the probability of being chosen while characteristics possess the large amount of information,the correlation coefficients between features is appended to the BPSO as a feature correlation factor determining the particle position.The experimental results show that the proposed method has better recognition performance by using the classifiers of J48,RF,KNN,MLP,NB and SVM.Finally,this thesis proposes a multi-sensor data fusion method based on weighted Linear Discriminant Analysis(LDA)in order to solve the question that a single type or position sensor cannot roundly accurately recognize the daily activities.For a given activity class,the problem is regarded as a binary classification problem and a data fusion model is constructed.The sensor weight vector is introduced to describe that sensors on different body positions may play as"experts" on different activity classes.Experiments indicate that this method can make a greater contribution to multi-sensor data fusion and better recognize daily activities.