| Summary: | 碩士 === 國立高雄應用科技大學 === 電機工程系博碩士班 === 104 === In this thesis, we propose a systematic calibration method to handle device diversity problem; a problem arises from the difference of wireless sensing ability among devices. Unlike conventional methods that rely on human labeled data or Wi-Fi landmark data, our method automatically generates pairwise points for calibration based on the RSS(Received Signal Strength) histogram. Conventional human labeled data is expensive and not convenient in particular applications. Recently, an automatically process has been proposed by using the Wi-Fi landmark concept. However, this framework requires a user to go through the environment to collect landmark data. This leads to a limitation that a user needs to wait for a whole until enough Wi-Fi landmarks are collected. To relax the limitation, we propose a method that can collect pairwise points automatically. First, we reach a room-level or region-level localization result by using relative features under the device diversity condition. Next, by calculating the Empirical RSS Cumulative Distribution Function (ECDF) of two devices in the same region, the RSS pairs for device calibration could be generated efficiently. Hence, the user does not need to wait for the collection of Wi-Fi landmarks. As a consequence, the proposed system could be suitable for particular application.
In addition, we also discuss a positioning solution in a wide region scenario. Practically, many Aps (Access Points) must be deployed to cover a wide region by Wi-Fi signal. Unfortunately, using more APs does not always provide a better positioning result. Where an AP is far away from a user may provide less robust information for localization. Therefore a method to select suitable APs is necessary to enhance the performance of the system. In summary, we proposed a two-stage process for localization in a wide region scenario. First, we use a coarse localization process to identify a region; second, some features (AP) are selected to active a fine positioning in each region. To active a region-level position, we use relative features with a neural network classifier since relative features are less sensitive to device diversity phenomenon. After the region is identified, we compare different AP selection methods to find out a suitable feature set for each region. Our experimental results prove that based on Principal Components Analysis (PCA) our system can reduce the effect from signal noise and keep the main structure of the original feature space. This leads to a better performance.
|
|---|
