Complex Activity Recognition In Multi-occupant Indoor Environments

Recognition of daily activities in the indoor environments is a key technology to discover the behavior intention,living habit,abnormal behavior and health status.This technology has been widely used in pervasive computing,context awareness,psychology,sociology,smart medical,smart living assistant,smart healthcare and so on.With the development of sensor technology and the enhancement of public privacy awareness,adopting indoor activity recognition with sensors is a hotspot of research.While,comparing with the direct-type sensors,like camera and wearable sensors,ambient sensors have various data kinds and not intuitive which led to greater difficulties in processing,recognition and management.The recognition scheme with ambient sensors puts up obvious problems.This paper mainly presents a complete generic recognition scheme,consists of sensor deployment and data transmit method,designing activity recognition algorithms for multi-complex environments.Selecting 14 typical activities in 4 spaces(kitchen,dining room,living room and bathroom)based on the activity features in multi-complex environments.Time Sequence Markov Logic Network(TSMLN)adopts the First-order-Logic(semantic modeling)with time sequence and Markov Network(statistical probability)which has good performance in data segmentation and activity recognition.Compared with existing Markov Logic Network Model,the accuracy of TSMLN is improved from 91%to 99.4%in interleaved situation,from 93%to 97.7%in concurrent situation.The main research aims of this paper is recognizing activities in the complex situations,especially for concurrent and interleaved activities in multi-occupant indoor environments,including activity recognition and user recognition.Noisy and data-missing is an unavoidable problem in recognition,distinguishing two similar activities is difficult in activity recognition,recognizing the new user is a typical problem in user recognition,the details have been shown in following:1)Solving the data-missing and noise disturbing problems by simplifying rules of TSMLLN.The situation is unavoidable,but more obvious in ambient sensor environment.In order to avoid the disturbance with a point missing or surplusage,sampling the rules to a simple chain with one action and one activity which can decrease the disturbance between actions and the dependence for knowledge expression,and improve the robustness.Because of the dependence between missing data or noise with real data,choosing Gibbs sampling is more accurate than MC-SAT(common method which is easy to fall into local optimum).Compared with TMLN in setting experiments,the accuracy of MLN with simplifying method is improved from 8.7%to 84%in data-missing situation,from 52.2%to 60.2%in noise disturbing situation.Compared with MC-SAT,the accuracy of Gibbs sampling is improved from 75.9%to 84%in data-missing situation,from 37.9%to 60.2%in noise disturbing situation.2)Distinguishing the similar activities by adding high-dimension time features and hierarchical structure to Markov Logic Network in activity recognition part.While the simple data of ambient sensor is limited.With the increasing of complexity in activity,the whole scheme for recognition will be overfitting.When the similar activities have been collected in a time window,the close degree(duration)and special period(period)features can divide the two similar activities better.Because of the dependence between data in two similar activities,I choose Gibbs sampling to inference the results.Compared with TSMLN in setting experiments,the accuracy of TMLN with high-dimension time features and hierarchical structure is improved from 23.6%to 92.6%in similar activity situation.Compared with MC-SAT,the accuracy of Gibbs sampling is improved from 68.6%to 926.6%.In addition,except the common sense,similar activities have the similar actions,but not in the same category,I have defined the new category which improve the generalization and decrease the complexity in processing.3)Presenting the occupant type recognition scheme based on user labels to realizing the new user recognition in user recognition part.The basic solution of existing research is establishing the specific model for every occupant which is a huge workload.This solution is not suitable for new commers.New commers will face the "cold start" problems.Occupant type is adopting the multiple combination method to discover the relationship with occupant preference,including time sequence preference,duration and period preference,location preference.Compared with TSMLN in setting experiments,the accuracy of TSMLN with user labels is improved from 8.7%to 90.3%by time sequence preference,from 8.7%to 84.9%based on high-dimension time(duration and period)preference,from 8.7%to 74.8%by location preference,from 8.7%to 98%by combination preferences.Finally,verifying the solution in three-occupant home,the accuracy of activity recognition is 98.3%and the accuracy of occupant type recognition is 84.7%which are proved that the scheme has practical significance and value in the environment of multi-occupant home.