Multi-user Indoor Activity Recognition Using Non-wearable Sensors

Over the past decades,human activity recognition(HAR)has been a significant branch in the field of computer science.With the advents of Internet of Things(Io T)era,HAR using sensors is gaining broader attention.Wearable sensor-based activity recognition is popular in mobile computing field,while non-wearable sensors are more suitable for intelligent environment applications,and this is the key to design and implement smart home.Furthermore,considering the convenience and non-disturbance of the equipment,non-wearable sensor-based activity recognition obtains growing attention.However,until now,activity recognition based on non-wearable sensors is still highly challenging,such as the generation of real data or the recognition results.Therefore,this paper studies the multi-user indoor activity recognition based on non-wearable sensors,in which non-wearable sensors include environmental sensors and object sensors.Indoor activities include some common daily activities,which may have an impact on the environment or the corresponding objects,including: studying in seat,boiling water,making coffee,talking and printing,reading,writing,programing,drinking.Therefore,the research of this paper can be divided into two parts according to the influence of the activities,the activity recognition research based on the environment sensor and the activity recognition research based on the object sensor.In the part of activity recognition research based on the environment sensor,power meters,temperature and humidity sensors,infrared induction sensor,obstacle avoidance sensor,light sensor and recording pen,are deployed to set up a simulation experiment platform in our office room.And we design and implement data collection system for daily activity data collection.In this paper,two algorithms and models are given for the recognition of activities based on the different data patterns and the characteristics of multi-user concurrent activities,the dynamic time wrapping(DTW)-based k-nearest neighbor(KNN)model and the multi-classifier model based on support vector Machine(SVM)algorithm.Finally,the multi-classifier model based on SVM algorithm is better,and 89% recognition accuracy is obtained,even if the results 74% of the DTW-based KNN model is also acceptable.In the part of activity recognition research based on the object sensor,the main sensors are the RFID tags and RFID reader.And there are two kinds of label,which are attached to the surface of each object.RFID reader is used to read the various received signal strength(RSS)data.Simulation platform is build,and further realize the RSS data collection system.In this part,the multi-classifier model with high performance in activity recognition based on environmental sensor is used to identify 93% recognition accuracy.Furthermore,taking into account the characteristics of RSS data,this paper further proposes the data mapping resized method,and then uses the CNN-LSTM model to extract the data space feature and time feature automatically,and finally achieves 95% recognition accuracy.At the end of this paper,a complete evaluation and comparative analysis of the activity recognition methods based on the environmental sensors and object sensors are carried out.The method of activity recognition includes the simulation environment construction strategy,the data acquisition process,and the activity recognition algorithm model.The overall design scheme of activity recognition based on non-wearable sensors is given.In conclusion,this paper presents a multi-user indoor activity recognition method based on non-wearable sensors,which achieves acceptable and superior recognition results better than the previous studies and has good scalability and portability.