| All of the wireless fidelity(Wi-Fi)localization,pedestrian dead reckoning(PDR)and Wi-Fi/PDR hybrid positioning based on smartphone are the hotspot of international and domestic research,as well as the most pervasive and widespread indoor positioning technologies.There are still a series of bottlenecks for them.For example,the fast construction of Wi-Fi fingerprint database,fast and accurate recognition for complex activities,accurate orientation estimation under complex activities,and good hybrid positioning strategy,et al.Study on the optimization problem of Wi-Fi/PDR hybrid indoor positioning method contributes to improve the usability and stability of indoor positioning system.It can not only enrich and improve indoor multi-source hybrid positioning methods and theries,but also provide technical support for indoor multi-source hybrid positioning.By taking the smartphone as research carrier and treating machine learning methods and optimization theory as the basic,distance similarity was systematically studied through theoretical research,simulation calculation and field test,as well as the fast construction of fingerprint database,the adaptive fingerprinting localization,the fast and accurate recognition of complex activities,the improved PDR considering activity recognition,and the hybrid positioning optimization method.The main contributions are as follows.(1)The Wi-Fi signal loss phenomenon was revealed.For this kind of phenomenon and the traversal mode between reference fingerprint and test fingerprint,the improved signal-domain distance and generalized hybrid distance were proposed.The proposed generalized hybrid distance can promote cluster results,and the improved signal-domain distance can improve the accuracy of clustering recognition and localization,which were verified through several experiments.Compared with the signal-domain distance using-100 dBm to replace the misssing signal value,the positioning accuracy was improved by 22.6% to 38% when there was no clusters,and the positioning accuracy was improved by 2% to 22.8%.To some extent,they showed the effect of signal misssing for fingerprint localization.(2)The fingerprint database construction method based on static crowdsourcing and adaptive path loss model interpolation was proposed after analyzing the advantages and disadvantages of dynamic or static fingerprints.Compared with the complete manual fingerprint database,inverse distance weighted(IDW)and Kriging interpolation fingerprint database,the proposed method got higher localization accuracy.And only less than 15.4% of the reference point data were used for constructing fingerprint database with the same positioning accuracy as the complete manual one,saving 85% time and effort.At the same time,the proposed method in this dissertation can also be used for rapid update of fingerprint database.(3)According to dynamic signal change,Wi-Fi fingerprint information not deeply mining,prone positioning error with fixed K value,clustering analysis were conducted based on generalized hybrid distance and affinity propogation clustering(APC)during offline stage,and then clusters were improved by considering transition regions.The adaptive weighted K nearest neiborhood(AWKNN)method based on location-domain and APC.The AWKNN positioning tests were conducted by using improved fingerprint union signal-domain distance,the positioning mean error(ME)was about 2.4 meters,and the root mean square error(RMSE)was about 1.9 meters.Compared with WKNN method which used the traversal reference fingerprint summation signal-domain distance,the ME was decreased by 1.4 meters,and the RMSE was decreased by 1.53 meters,with the positioning accuracy improving by 37%.And the stability was improved a lot.At the same time,the proposed method in this dissertation weakened the influence of K value selection on positioning accuracy and avoided repeated evaluation of positioning accuracy with different K values,achieving more accurate localization by setting a large K value.(4)Aiming at the problems of poor real-time performance of existing activity identification research,extensive use of frequency-domain characteristics and not considering indoor positioning interference activities,the common complex activities of indoor pedestrians were subdivided and defined based on immense analysis of sensor data.The vertical and horizontal shaking were introduced into indoor activity recognition.The static and movement activities were subdivided to reduce the disturbance of indoor positioning caused by standing still,shaking and other activities.The basis of the time window selection considering real-time performance of activity recognition was established.And the cosine differennces of orientation angle were used for activity recognition.In order to avoid a large number of frequency calculations,a method of identifying indoor complex activities only using time domain features based on random forest(RF)was proposed.By comparing and analyzing the classification results of naive bayes,decision tree,support vector machine,RF and neural network classification algorithm,it was found that the extracted time-domain features proposed in this dissertation had a high accuracy for both individualized and universal activities recognition.The accuracy of individualized and universal activity recognition based on RF were 99.78% and 99.5%,respectively.Experimental results showed that the proposed method in this dissertation can be used for rapid and accurate recognition of indoor complex activities.(5)Aiming at the problems of weak anti-interference ability,poor real-time performance and inaccurate oreientation estimation of PDR,the change of sensor data under different activities was analyzed emphatically.And gait detection methods considering time synchronization under different activities,step estimation model based on random forest regression and oreintation estimation methods with waveform calibration were proposed,forming the system of the improved PDR methods considering activity recognition.Experiments of PDR were carried out under walking with smartphone holding and walking with the smartphone in the pocket,respectively.The walking distance was about 211 meters.Gait detection,step length estimation and heading estimation were performed for these two activities.The accuracy of steps counting under two activities were up to 99.6%,the accumulative errors of step length estimation were less than 2.2 meters,and errors of orientation estimation were less than 6 degrees.Closeure errors were less than 2.4 meters,the location divergence rates were less than 1.2%,the positioning MEs were less than 1.8 meters,and the RMSEs were less than 1.4 meters.Experimental results showed that the improved PDR methods considering activity recognition had strong anti-interference ability,accurate heading estimation and high positioning accuracy.(6)Aiming at the problems of largely changing Wi-Fi fingerprint localization results with low update frequency,error-prone clustering recognition,and the rebound phenomenon of hybrid localization results,the Wi-Fi/PDR hybrid localization optimization method based on position constraint,displacement constraint and direction constraint was proposed.Hybrid localization optimization experiments were respectively carried out under two activities,walking with smartphone holding(WH)and walking with the smartphone in the pocket(WP).The positioning results under these two activities had high coincidences with the real trajectory.The positioning ME under WH activity was less than 1.6 meters,and the RMSE was about 1 meter.The positioning ME under WP activity was about 1.62 meters,and the RMSE was about 1.4 meters.Compared with the hybrid positioning results based on Wi-Fi and PDR before optimization under WH activity,the positioning ME after optimization was decreased by 0.91 meters,the RMSE was decreased by more than 1.9 meters,with the positioning accuracy improving by 36.5%.Experimental results showed that the proposed method can greatly improve the accuracy of hybrid positioning and promote the performance of indoor positioning system,solving the problems of largely changing Wi-Fi fingerprint localization results with low update frequency,error-prone clustering recognition,avoiding the rebound phenomenon of hybrid localization results. |
PDF Full Text Request