Machine Learning Based Obstacle Detection for Automatic Train Pairing

TL;DR

This paper evaluates the use of supervised machine learning, specifically ensemble tree methods, with UWB devices for obstacle detection in train pairing and AGV scenarios, achieving around 95% accuracy.

Contribution

It introduces a novel approach of applying supervised learning to UWB-based obstacle detection, demonstrating high accuracy in simulation.

Findings

Ensemble tree methods achieve ~95% obstacle collision prediction accuracy.

UWB devices can be effectively used as both ranging radios and radars.

Supervised learning improves obstacle detection in UWB-based systems.

Abstract

Short Range wireless devices are becoming more and more popular for ubiquitous sensor and actuator connectivity in industrial communication scenarios. Apart from communication-only scenarios, there are also mission-critical use cases where the distance between the two communicating nodes needs to be determined precisely. Applications such as Automatic Guided Vehicles (AGV's), Automatic Train Pairing additionally require the devices to scan the environment and detect any potential humans/obstacles. Ultra-Wide Band (UWB) has emerged as a promising candidate for Real-Time Ranging and Localization (RTRL) due to advantages such as large channel capacity, better co-existence with legacy systems due to low transmit power, better performance in multipath environments etc. In this paper, we evaluate the performance of a UWB COTS device - TimeDomain P440 which can operate as a ranging radio and a monostatic radar simultaneously. To this end, we evaluate the possibility of using Supervised Learning based estimators for predicting the presence of obstacles by constructing a multiclass hypothesis. Simulation results show that the Ensemble tree based methods are able to calculate the likelihood of obstacle collision with accuracies close to 95%.