Study On Human Activity Recognition Method Based On Indoor Location And Multiple Contexts

The pervasive computing requires ubiquitous computers and sensors presenting context awareness and changing their behavior accordingly,taking the initiative to adapt to the needs of users and provide intelligent services.Location awareness is an important part in context awareness,and by using the global satellite navigation and positioning system,the positioning accuracy can reach meter-level in the outdoor.However,in the urban canyon and indoor,radio signal can be blocked and the transmission multi-path effect can be strong,thus the satellite systems have limitations.In this circumstance,on the one hand,more and more indoor positioning solutions have been presented,but the various types of indoor awareness technologies have pros and cons and they need fusion.On the other hand,smartphones not only have high performance in computation,but also have been integrated a wealth of sensors which provides a good hardware platform for fusion.Additionally,location-based services have been widely used in all aspects of human life,and it includes returning information to user when they querying and pushing services according to the needs of users.But in the latter way of service,the user activity analysis and the demand determination are only based on the user's location,and this method is not flexible thus the user experience is poor.In order to improve the location service experience and provide intelligent services,the preconditions are accurate recognition of human activities with multiple contexts including location and prediction of future trends in the activities.Focus on issues of location awareness about continuous high precision indoor positioning and elevation awareness especially the floor information,and based on smartphone platform and current indoor location awareness technologies,this paper conducted research on WiFi positioning,fusion of WiFi/IMU and floor information obtaining using barometer.Meanwhile,focus on the accurate recognition of human activities problem based on location awareness,this paper explored the multiple contexts method.Specific research contents and achievements are as follows:(1)Analyzed the channel propagation of WiFi RSSI.Based on the analysis of RSSI temporal and spatial patterns,the error components in the RSSI measurement were analyzed and the systematic deviation of the drift over time was put forward.The experimental results show that the method is feasible,and it improved the RMS of positioning error to about 4 meters no matter using deterministic or probabilistic algorithm,and the effect is obvious especially for deterministic algorithm.(2)The quality control of WiFi RSSI fingerprint positioning system was studied,and the impact factors such as access point density,access point distribution,propagation attenuation factor,environment noise and reference point density were studied in detail.Experiment results show that the high access point density,high propagation attenuation factor and low ambient noise are favorable to obtain accurate position,while the reference point density is more favorable for positioning at about 1 meter,but the distribution of access points has no influence.(3)For the problem that positioning results of WiFi RSSI is discontinuous,it proposed to use the smartphone IMU PDR to carry out fusion positioning.Considering the computing efficiency of the computing platform,the UKF fusion algorithm was used.In order to make the evaluation of the positioning results more comprehensive,an accurate evaluation method based on the accuracy of static and dynamic was used.By using the high precision GPS/INS as the ground truth reference system,the experimental results show that positioning is continuous and the average error of the fusion system is 3.52 m,which is less than individual WiFi system.(4)In view of the problem that the indoor positioning elevation reach is little,the atmosphere environment is similar in the same building,and the indoor elevation is utilized with the floor information.The paper presented the identification of the floors based on the differential pressure method.The experimental results show that the height difference error is less than 1 meter which is less than the common floors height 3 meters,thus the method can be used to determine the floor.(5)Aiming at the problem of activity recognition based on smartphone sensors,this paper proposed a method using feature extraction based on multiple contexts framework and the machine learning method.The multiple contexts include temporal context(time),spatial context(location),temporal and spatial context(location duration)and user context.The results show that the framework can effectively recognize the eight complex activities in campus activities,and the recognition accuracy can reach 84.1%.(6)There are a lot of algorithms in machine learning,thus this paper built a linear optimization model for multiple contexts activity recognition.The Weka platform was used to evaluate 18 kinds of offline learning algorithms,and the evaluation model was put forward.When the weights of classification accuracy,model learning efficiency and activity recognition efficiency is 1.0: 0.8: 1.2,the REPTree algorithm in the framework of decision tree has advantages in the comprehensive evaluation results.Aiming at the problem of limited accuracy of naive Bayesian learning in the original framework,and according to the evaluation result,this paper used the random forest algorithm to improve.The experimental results show that the weighted average F1 measurement of the activity recognition was greatly improved.In view of the shortcomings of low learning efficiency and cannot update the model in real time,this paper put forward the online learning to improve.The experimental results show that the online learning method is efficient and can use the new data effectively and update the model in real time.