Research On Daily Activity Segmentation And Feature Calculation Methods In Smart Home

The rapid development of the Internet of Things and sensor technology has brought smart homes into people’s lives.Daily activity recognition technology analyzes environmental data collected in smart homes to determine the daily activities that residents are performing,and then perceive residents’ needs.The smart home system can solve the problem of independent living of the elderly to a certain extent,help and improve the daily life of the elderly,and can reduce the cost of care.The problems in daily activity recognition are:(1)In the field of behavior segmentation,most researchers use sliding window technology,and do not use sensors that are closely related to activity as a segmentation basis.(2)Most researchers focus on the activity recognition of a single user in smart homes,and have not made too many attempts on multi-user activity recognition.(3)In the previous research methods,most of the first statistics of the frequency of sensor appearance,and then the frequency as a behavior feature for activity recognition.This method does not consider the influence of other users’ noise data on the recognition result in the multi-user activity recognition.This article has conducted in-depth research on the above-mentioned problems and achieved the following research results:The effect of daily activity recognition depends on the quality of sensor data stream segmentation.Although segmentation technology has made great progress in long-term research,the existing segmentation technology still divides an activity into multiple activities,or merges multiple activities into one activity.In response to this problem,this thesis proposes an activity segmentation method based on best-fit sensors to solve the problem that the segmentation boundary of dynamic sensor data stream is difficult to determine.This method extracts the activity starting sensor and the corresponding activity and the duration of the activity from the training set,and then calculates the information gain value of each sensor in each activity,obtains the best fitting sensor for each activity,and establishes a model to perform data flow.segmentation.In addition,daily activity recognition usually uses the trigger frequency of the sensor to identify the type of activity.However,in a multi-user smart home,there are overlapping and nesting behaviors between users.This makes different activities share the same sensor sequence,resulting in an increase in noise data in the data set.However,due to the different duration of the same activity,the trigger frequency of the sensor is different,which will cause the same feature to have a different degree of influence on the same activity,thereby affecting the activity recognition effect.In order to solve this problem,this thesis proposes a multi-user daily activity recognition method based on WFO(Weighted Frequency and Odds)feature calculation.This method eliminates the influence of noise data and sensor trigger frequency on activity recognition by assigning higher weight to sensors that are important to activities,and improves the accuracy of multi-user activity recognition.The experimental data in this article comes from the public data set of the Center for Advanced Research on Adaptive Systems(CASAS)at Washington State University.Aiming at the problem of activity segmentation,this thesis conducts experiments on two data sets with one user.The experimental results show that the activity segmentation method based on the best-fit sensor can effectively determine the boundary of the activity.Aiming at the problem of multi-user activity recognition,this thesis conducts experiments on a data set with two users.The experimental results show that the method proposed in this thesis can more effectively identify the activity of multiple users in smart homes.