Research On Learning Based WLAN Indoor Positioning Techniques

With the pervasively available of wireless networks and rapid development ofmobile smart devices, location based services (LBSs) have attracted more and moreattentions. LBSs have been widely used in many areas such as emergency rescue, healthcare, social network, navigation and monitoring, and shown broad market prospects.Accurate indoor positioning technique is a prerequisite and critical component to LBSs.Most of indoor positioning techniques require expensive hardware devices, andhave small application area, thus limiting the popularization of indoor LBSs. Amongthem, Received Signal Strength (RSS) based positioning in wireless local area network(WLAN) has been paid special care due to its cost effectiveness and relatively highaccuracy. WLAN positioning is entirely founded on the existing WLAN infrastructuresand mobile devices, no extra device is needed. However, it is a challenging task toimprove accuracy in the complex indoor radio propagation environment. This is becauseRSS shows highly uncertain patterns caused by various factors, such as mulit-patheffect, human absorption and signal interference, and so that renders a nonlinear andmany-to-many mapping relationship between RSS and physical locations.Researches on WLAN fingerprinting indoor positioning technique are taken in thisdissertation. Through analysis of the research status at home and abroad, problems anddrawbacks of existing algorithms are pointed out. These problems and drawbacks existin the main modules of positioning system, including location clustering, access pointselection, and location feature extraction. The goal of this dissertation is to reduceuncertainty of RSS and improve these algorithms, by adopting the latest theories andtechnologies in the machine learning area, including kernel based learning, manifoldlearning, information theory and support vector machines.Firstly, researches on location feature extraction from RSS signals are carried out, anda novel positioning algorithm (named LKDDA-APS) based on kernel directdiscriminant analysis (KDDA) is proposed. Due to the complex indoor radiopropagation environment, directly taking RSS signals as inputs to positioning algorithmmay introduce considerable redundant and noisy information, and thus deterioratingpositioning accuracy. Therefore, a feature extraction algorithm is required to fuse RSSsignals and discard redundant and noisy information. To address the problem thatprevious algorithms cannot adapt the nonlinearity of RSS well, KDDA for featureextraction is proposed. Besides, LKDDA-APS also incorporates with the proposedlocation clustering, access point (AP) selection and support vector regression (SVR) inthis dissertation. Experimental results show that, compared with previous algorithms, LKDDA-APS achieves significant accuracy improvement, while reducing calibrationeffort considerably.Secondly, to save energy consumption on mobile devices, a novel positioningalgorithm (named LLDE-APS) based on local discriminant embedding (LDE) isproposed. To protect privacy of the users, WLAN fingerprinting technique tends todeploy client-based architecture, which performs all computation process on mobiledevices. The client-based architecture requires positioning algorithm to reduce energyconsumption on mobile devices as much as possible in order to lengthen the battery lifeof the mobile devices. Hence, this dissertation proposes LLDE-APS based onLKDDA-APS. Instead of KDDA, LDE with much lower computation cost is proposedto extract low dimensional features from high dimensional RSS signal space.Experimental results show that, without significant accuracy deterioration thanLKDDA-APS, LLDE-APS improves accuracy significantly than previous algorithms.Moreover, LLDE-APS reduces the online computation cost than LKDDA-APS, thussaving much energy consumption on mobile devices.Thirdly, researches on the selection of the source of RSS signals are carried out, and ajoint AP selection scheme is proposed. Not all APs are beneficial for accuracyimprovement, and different APs contain different location information quantity. Hence,the discrimination power of each AP or APs subset should be measured, and the bestAPs subset is selected. Unlike previous AP selection schemes never considering thecorrelation among RSS signals, the proposed joint AP selection scheme jointly selectsthe optimal APs subset with the maximum mutual information. Experimental resultsshow that, the proposed AP selection scheme performs better than previous schemes,thus improving positioning accuracy and reducing computation cost significantly.Finally, researches on the location clustering for positioning region are carried out,and a novel location clustering algorithm is proposed. In large-scale positioning region,the positioning model founded on the whole positioning region is sub-optimal, becausethe statistical properties of RSS signals vary with physical locations. Hence, it isnecessary to constraint the positioning model into sub-regions through locationclustering algorithms. To ensure high classification accuracy of location clustering, theproposed location clustering algorithm incorporates k-means clustering with supportvector classifier. Experimental results show that the proposed location clusteringalgorithm obtains higher classification accuracy than previous algorithms.