DeepMTL Pro: Deep Learning Based MultipleTransmitter Localization and Power Estimation
In this paper, we address the problem of Multiple Transmitter Localization (MTL). MTL is to determine the locations of potential multiple transmitters in a field, based on readings from a distributed set of sensors. In contrast to the widely studied single transmitter localization problem, the MTL problem has only been studied recently in a few works. MTL is of great significance in many applications wherein intruders may be present. E.g., in shared spectrum systems, detection of unauthorized transmitters and estimating their power are imperative to efficient utilization of the shared spectrum. In this paper, we present DeepMTL, a novel deep-learning approach to address the MTL problem. In particular, we frame MTL as a sequence of two steps, each of which is a computer vision problem: image-to-image translation and object detection. The first step of mage-to-image translation essentially maps an input image representing sensor readings to an image representing the distribution of transmitter locations, and the second object detection step derives precise locations of transmitters from the image of transmitter distributions. For the first step, we design our learning model Sen2Peak, while for the second step, we customize a state-of-the-art object detection model YOLO-cust. Using DeepMTL as a building block, we also develop techniques to estimate transmit power of the localized transmitters. We demonstrate the effectiveness of our approach via extensive large-scale simulations and show that our approach outperforms the previous approaches significantly (by 50% or more) in performance metrics including localization error, miss rate, and false alarm rate. Our method also incurs a very small latency. We evaluate our techniques over a small-scale area with real testbed data and the testbed results align with the simulation results.