Identifying the positions of mobile devices within indoor environments allows for the development of advanced context-based applications and general environmental awareness. Classic localization methods require GPS; an expensive, high power consuming and inaccurate solution for indoor scenarios. Relative positioning instead allows nodes to recognize location in relation to neighboring nodes without the requirement of GPS. To triangulate their own position however, indoor localization methods either use Received Signal Strength Indication (RSSI) retrieved from neighboring devices to determine distance or simple binary contact information denoting whether two nodes are in communication range of one another. RSSI however is plagued by many sources of noise, thus decreasing distance prediction accuracy as well as being unreliable for networks of heterogeneous devices. Further, using only binary contacts provides a limited information for localization. In our work, we first demonstrate the unreliable nature of RSSI in heterogeneous networks. We then demonstrate our intermediate solution between unreliable RSSI and oversimplified binary classifications by introducing Perceived Direction Information (PDI) composed of three states: approaching, retreating and invisible. Through real world experiments, we demonstrate that PDI can be predicted using a Dense Neural Network with more than 95% accuracy even on devices not used during training. We then describe an anchorless Monte Carlo Localization (MCL) algorithm which uses PDI to achieve higher accuracy and a reduction of communication over the state-of-the-art MCL based methods.

识别室内环境中移动设备的位置可以开发高级的基于上下文的应用程序和一般的环境意识。经典的定位方法需要GPS；针对室内场景的昂贵，高功耗且不准确的解决方案。相对定位允许节点识别相对于相邻节点的位置，而无需GPS。然而，为了对自己的位置进行三角测量，室内定位方法要么使用从相邻设备中获取的接收信号强度指示（RSSI）来确定距离，要么使用简单的二进制联系信息来表示两个节点是否处于彼此的通信范围内。但是RSSI受到许多噪声源的困扰，因此降低了距离预测精度，并且对于异构设备的网络也不可靠。此外，仅使用二进制触点会为本地化提供有限的信息。在我们的工作中，我们首先证明异构网络中RSSI的不可靠特性。然后，我们通过介绍由以下三种状态组成的感知方向信息（PDI）来展示不可靠的RSSI和过于简化的二进制分类之间的中间解决方案：接近，后退和不可见。通过现实世界的实验，我们证明可以使用具有超过95的密集神经网络来预测PDI 然后，我们通过介绍由以下三种状态组成的感知方向信息（PDI）来演示不可靠的RSSI和过于简化的二进制分类之间的中间解决方案：接近，后退和不可见。通过现实世界的实验，我们证明可以使用具有超过95的密集神经网络来预测PDI 然后，我们通过介绍由以下三种状态组成的感知方向信息（PDI）来展示不可靠的RSSI和过于简化的二进制分类之间的中间解决方案：接近，后退和不可见。通过现实世界的实验，我们证明可以使用具有超过95的密集神经网络来预测PDI即使在训练过程中未使用的设备上，准确度也达到％。然后，我们描述了一种无锚点的蒙特卡洛定位（MCL）算法，该算法使用PDI来实现更高的准确性，并通过基于最新MCL的方法来减少通信量。