| dc.description.abstract | Indoor positioning still faces important challenges from its current mainstream contenders, to support accurate and energy-efficient positioning. Inertial sensors-based solutions have cumulative errors and drift issues, camera-based schemes have privacy issues, and feature matching is affected by different views (of the same object). Wi-Fi-based positioning is widespread indoors, but its high energy consumption problem limits its application in receiver mobile devices such as smartphones. To improve the performance of the above three indoor localization schemes, an Efficient Neural Inertial Localization (ENILoc) for human position estimation is proposed by considering different waypoints to revise the cumulative errors and drifts. First, a more minimal model that only consists of simple Linear layers and Rectified Linear Unit (ReLU) activation functions, and these waypoints are detected by human activity recognition (HAR) by creating a corresponding relationship between both of them. Then, in addition, a Fast Image-Based Indoor Localization Using an Anchor Control Network System (FILNet) is developed, where framed pictures hanging on the wall are utilized as the key information to create an image database with position information and then, infer the user’s location by affine stability algorithm and fast spatial-indexing. Furthermore, a Wi-Fi Localization with Deep Spiking Neural Network (SNN) is introduced, called SpikeWL, which is a Wi-Fi fingerprint-based multi-floor localization scheme, the Wi-Fi received signal strength indication (RSSI) is employed to perform buildings and floors recognition (SpikeWL-BFR) and Indoor Position Regression (SpikeWL-IPR). The main contributions of this research are as follows. First, ENILoc utilizes a simple linear layer to challenge the booming Transformers or Convolutional layer-based long-term human position estimation, greatly improving its training speed (2.5-6x) and decreasing memory usage (17-31x) while achieving better or similar localization errors. Second, FILNet uses framed pictures to create the image fingerprint database, which enhances privacy protection. Additionally, the proposed multi-view supplementary algorithm and fast spatial indexing reduce localization errors and improve the recall of feature matching. Third, a SpikeWL is proposed that adopts SNN to classify different buildings and floors, and then, to predict users’ positions using the Wi-Fi Received Signal Strength Indicator (RSSI). Compared to Artificial Neural Network (ANN) based Wi-Fi Localization, this approach reduces energy consumption by over 90% while enhancing the classification accuracy of buildings and floors and reducing localization errors. | en_US |
