The idea of adding sensors and intelligence to basic objects has been around since the 1970s. The Internet of Things was initially most interesting to business and manufacturing but with the lowering cost of processors and advancements made in wireless networks and data analytics, the emphasis is now on implementing IoT frameworks in a wider range of applications. A large number of these applications requires tracking capabilities for outdoor localization and is therefore dependent on GPS. The problem with using GPS for IoT applications is the battery lifetime and what that entails in terms of cost and convenience. Low Power wide area networking (LPWAN) technologies represent a very good alternative to GPS. LoRa in particular has a low energy consumption and a long range due to the good sensitivity of the receivers.
Here at the Swisscom Digital Lab in the scope of my master thesis, I am developing a LoRa based localization technique to pinpoint the position of a moving subject in real time. I am particularly interested in urban areas with a high density of buildings where the signal strength becomes increasingly noisy due to non-line of sight. There are multiple use cases for this technology such as tracking packages or containers in a storage yard, finding a car location in a large parking lot or monitoring the flow of goods in supply chains.
LoRaWAN messages suffer from both temporal and spatial non-stationarity which puts standard trilateration techniques at disadvantage. The idea is to circumvent the noisiness of the signal and build an algorithm that can adapt to the environment. The setup consists of 5 office gateways put in different buildings, 1 LoRa module moving in a random walk fashion and GPS as ground truth measurements. Using deep learning techniques and taking into account the timeseries characteristic of datapoints, a mean error localization of 20 m was achieved for a person walking in the EPFL innovation park. The results for static localization are even more promising.
As you can see in the animation above, we can track the position (in red) online by giving a prediction (in green) without any lag. Depending on the situation, If we restrict the model’s estimation to only predefined paths the accuracy becomes higher.