Machine learning based sensor fusion for localization applications

Machine learning based sensor fusion for localization applications Introduction Facts Research Solution Results Accuracy: >75% Firefighters usually face harsh conditions and have to respond to emergencies where life is at risk. Satellite-based localization increases rescue workers’ safety. However, its reliability and accuracy can be poor indoors. This research was conducted in the context of the Autonomous Indoor & Outdoor Safety Tracking System (AIOSAT) H2020 European project, which aims to overcome the existing limitations of Global Navigation Satellite System (GNSS) by fusing sensory data of Pedestrian Dead Reckoning (PDR), Ultra-Wideband (UWB). FIRES IN EUROPE PER YEAR: 2 million fires 50.000 injuries 5000 deaths 10-20 firefighter The traditional approach for de-noising sensor data and predicting future values is usually based on sequential Monte Carlo methods such as particle filters. This research aims at filtering out the noise and predicting future values for sensor data by investigating the performance of Neural Networks and implementing sensor fusion based on it. Recurrent Neural Networks are especially suited for time series based data because they remember information about the past states and can use it when processing new data. Convolutional Neural Networks are very effective for deriving high-order features out of a fixed-length segment of an overall dataset and work particularly good for anlysis of a time series of sensor data. Due to the fact that these 2 types of neural networks are highly efficient and effective in working with sensor data, it was decided to combine them for achieving a better accuracy. In general, neural networks are highly sensitive to the values of their hyperparameters. For the purpose of hyperparameter tuning, Keras Tuner was used. Most of the work in this research has been performed on simulated data. Firstly, data has been generated for latitude and longitude, based on the GNSS sensor. For simplicity, the training started with straight line paths and finding a proper model for the purpose of denoising such paths. An example of de-noising a straight path through Neural Networks prediction can be seen in the figure bellow. The same techniques have been applied for more complex path types, consisting of a combination of straight and curved line seg - ments. In the figure on the right an example of a prediction for a curved path is shown. Accuracy: > 97% Positioning Error: < 10 cm The sensor fusion implementation is a work-inprogress. The data consists of signals generated bythe GNSS and PDR sensor. The main challenge lies in the irregular, non-continous charac - ter of the signals. There is no PDR and GNSS signal at the same time for every time step. A concept model architecture is proposed for this particular situa - tion. The system supports an ad-hoc configurable rescue workers deployment. Its decentralized architecture relies on local data-processing nodes arranged in hierarchical clusters. Communication between the nodes and clusters is facilitated by multiple channels (LoRaWAN, BT, LTE) that provide high reliability through various fall-back strategies supplemented by decision logic, A reliable application backend allows the brigades and mission commanders to track the location of the rescue workers during a mission. Objective Tania-Andreea Grama*, Dawid R. Zalewski, Javier Ferreira-Gonzalez Ambient Intelligence Research Group, Saxion University of Applied Sciences, M.H. Tromplaan 28, 7513 AB Enschede * corresponing author, taniag[email protected]