| In many real sceneries, people have strong desire of indoor localization, such as thepurchase localization navigation in a mall, entrance, toilet and shop of huge airpots andrailway stations. However, indoor localization techniques have not been put into businessas outdoor localization techniques, such as GPS and Telecom base station, the reasons are,signal sources have a long distance to the buildings, and people keep moving indoor, andthe indoor environment is quite complicated, signals’ strength decreases because of reflect.According to the restricted eyesight indoor, indoor localization needs the accuracy scalesnearer than that of the outdoor localization, but the current applied indoor localizationtechniques can hardly solve these problems and thus only exists in researchenvironments.We are now in a large data era. Pattern recognition techniques, especia llystatistics pattern recognition techniques are developed. Here, we apply the patternclassification technique to transfer the indoor localization problem into a patternclassification problem.Resent days, Wi-Fi is the most popular many-to-many wireless network, we can usepattern recognition, especially statistic pattern recognition techniques, to distinguishdifferent position fingerprint of different addresses. This paper applies pattern recognitiontechniques, to transform the indoor localization problem into a classifying problem, inorder to complete performing indoor localization.In this paper, we have done some new works: we implement the whole indoorlocalization system resolution, which consists of three parts, the signal collect part, themodel training part, and the indoor localization part. We observe that it’s impossible toprevent a classifier from classifying wrongly, so we adopt theory of probability to provetwo questions:(1) If each classifier has an accuracy of more than50%, the correct rate of manyclassifiers using majority voting rule shall be higher than that of a single classifier.(2) The more classifiers we use, the correct rate of many classifiers using majorityvoting rule shall be much more higher than that of a single classifier.We hava completed designing and building this system, and proved the twoconclusions above. Finally, the accuracy of our system is nearly the same as that ofUniversity of California at Berkeley. |
