Seite - 32 - in Joint Austrian Computer Vision and Robotics Workshop 2020

4.1.Weighting Firstan intensityestimation αˆm foracertainmea- surementm is calculatedbasedon thesingleparticle p∈P tobeweightedandthe influenceofallalready definedsources sˆ∈Ψassumingthemodelpresented inEquation 1: αˆ(m) =αbgr+ αp 4 ·pi ·dp(m)2 + ∑ sˆ∈Ψ αs 4 ·pi ·ds(m)2 (2) UsingEquation2 the relativemeansquareerrorcon- sidering allmeasurements is calculated like: ermse= 1 |Γ| ∑ m∈Γ ( αˆm−αm αˆm )2 (3) where |Γ| is the number of measurements. The weight for a singleparticle is thencalculated like: wp= 1 1+ermse (4) After all particle weights have been updated the weights are normalizedsuch that ∑ p∈Pwp= 1. 4.2.Re-SamplingandClustering During re-sampling a certain percentage of parti- cles with the highest weights stay the same, while another percentage of particles with the smallest weightsareomittedandnewlydrawnfromauniform distribution over the search space. The remaining particles are re-sampled by adding Gaussian Noise to the intensityαp and position based on the parti- clesweight: [x′p,y ′ p,α ′ p] T∼N ( [xp,yp,αp] T, diag(σpos,σpos,σint) 1+wp ) (5) The total number of particles stays the same. After a defined number of iterationsk the particles are clus- tered using the mean shift algorithm as suggested by [3]. The cluster centroids have the same structure as asingleparticleandarethenevaluatedbytheweight- ing algorithm described in section 4.1. If the weight of a cluster surpasses a defined thresholdϕ the cen- troid is believed to be a real source and added to the growing setofpredictedsourcesΨ. 5.ExperimentalEvaluation -ENRICH2019 Aspartof theTUGrazRoboticsTeamTEDUSAR weparticipated in theEuropeanRoboticsHackathon TotalParticles 2000 RandomNewParticles 10% SustainParticles 10% MaxIterationsT 1000 ConfidenceThresholdϕ 0.9 Clustering Intervalk 20 iter PositionDeviationσpos 0.5 IntensityDeviationσint 0.4 Table 1: Hyper-parameters used for the ENRICH 2019 Figure 1: Source estimation based on live measure- mentdataduring theENRICH2019inZwentendorf. - ENRICH 2019 at the nuclear power plant Zwen- tendorf, Austria1 and were able to test our particle filter approach under real world conditions. An au- tonomous robot created a 3D map of the interior while our approach created the mathematical model of the real radiation sources and the radiation con- tamination. The parameters used are shown in Table 1. An experimental result can be observed in Figure 1. In this experiment two sources were placed in a largerroom. After traversingtheroomandcollecting radiation measurements our approach correctly pre- dicted the locationand intensityof the twosources. 6.ConclusionandFutureWork In thispaperwepresented theadaptationofanap- proachbasedonaparticlefiltertodeterminetheloca- tionandintensity foranarbitraryandunknownnum- ber of stationary radiation sources. This approach has been successfully tested and proven to be appli- cable in real world scenarios, like an accident in a nuclear facility. Future work will focus on reducing thenumberofhyper-parameters. 1www.enrich.european-robotics.eu 32