Seite - 31 - in Joint Austrian Computer Vision and Robotics Workshop 2020

UGVRadiationMappingusingaParticleFilter AlexanderPermann,DanielHettegger,GeraldSteinbauer InstituteofSoftwareTechnology,TUGraz alexander.permann@student.tugraz.at, daniel.hettegger@alumni.tugraz.at steinbauer@ist.tugraz.at Abstract. We present and evaluate a particle fil- ter based approach to predict the location and emis- sionintensityofanarbitraryandunknownnumberof stationary nuclear radiation sources from measure- ment data taken by an autonomously navigating un- manned ground vehicle (UGV). 1. Introduction Due to the threat for humans caused by radiation and the associated difficulties after a nuclear disas- ter it is crucial to establish save methods of estimat- ing the radiation distribution in certain affected ar- eas. For this purpose we suggest to record radia- tionmeasurementusinganautonomousUGV.These measurements are then processed by an adapted par- ticle filter to generate a detailed radiation distribu- tion model of the affected area. The approach pre- sented in this paper has been successfully tested in realisticconditionsat theENRICH2019—European Robotics Hackathon, where live radiation sources had to be detected inside the nuclear power plant in Zwentendorf,Austria. 2.RelatedResearch In [1] Eric T. Brewer used an autonomously fly- ing aerial platform to detect and locate a single ra- dioactive point source using a particle filter. In [2] M. Morelande et al. compare the performances of a maximum likelihood estimator and a Bayesian esti- mator approach to deal with an unknown number of sources. D. Shah et al. present a particle filter in [3] thatmanages to locatemultiple radiationsources. 3.ProblemDescription The setting is represented by a set Θ of unknown radiation sourcessand a setΓof radiation measure- mentsm. Thegoal is togenerateasetΨofestimated sources sˆ, that fits the number and intensities of the real sources accurately. Each set holds elements de- finedbyacertainlocationxiandyiandanequivalent radiationdose rateαi inSvs−1 that either represents the actual measurement for the set Γ or the theoret- ical dose rate that would be measured at the exact position of a source for the sets Θ and Ψ. In general formodellingtheradiationintensityatacertain loca- tion lbasedonasetofsourcesΘ,weassumethat the radiation follows the principle of superposition and the inverse-square-law which has been shown to be applicablebymultiple formerapproaches. [1,2]: α(l) =αbgr+ ∑ s∈Θ αs 4 ·pi ·ds(l)2 (1) whereαbgr denotes the known background radiation andds(l) theeuclideandistancebetweenthelocation l and the sources. 4.ParticleFilter In contrast to common particle filter use cases in robotics (e.g., estimatinga robotsposition), it isnow necessary todetectmultiplesources thatcanco-exist at thesametimeatdifferentpositions. In thiscontext particlesarepredictionsofpotential sources [3]with each particlep∈P being represented similar to real sourcesbyp=<xp,yp,αp,wp>withanadditional weightwp that is related to theprobability that acer- tain particle has the parameters of a real source. At first theparticlesare initializeduniformlydistributed on theplanewhere themeasurements tookplaceand givenarandomintensitywithinthesamerangeofthe measurement results. The algorithm then iteratively performs the twostepsofweightingand re-sampling and adds estimated sources sˆ to the growing set Ψ until a maximum number of iterationsT is reached andΨ representsaconsistent estimation forΘbased on themeasurementsΓ. 31