Integrating SLAM into Gas Distribution Mapping Achim Lilienthal, Amy Loutfi AASS, Dept. of Technology, Örebro University Jose Luis Blanco, Cipriano Galindo and Javier Gonzalez System Engineering & Automation Dept., University of Malaga
→ Contents Gas Distribution Mapping Achim J. Lilienthal
Gas Distribution Mapping Achim J. Lilienthal
Localization (SLAM) Achim J. Lilienthal
Localization (SLAM) Achim J. Lilienthal
Gas Distribution Mapping + SLAM + position, position uncertainty Achim J. Lilienthal
Contents 1. Applications of Gas Distribution Modelling? 2. The Challenges for Gas Distribution Mapping 3. Kernel Based Gas Distribution Mapping 4. Integrating SLAM and Gas Distribution Mapping 5. Experiments and Results 6. Summary Achim J. Lilienthal
→ Contents Applications of Gas Distribution Modelling Achim J. Lilienthal
1 Gas Distribution Modelling Applications Oil Refinery Surveillance Achim J. Lilienthal
1 Gas Distribution Modelling Applications Oil Refinery Surveillance Garbage Dump Site Surveillance Achim J. Lilienthal
1 Gas Distribution Modelling Applications Oil Refinery Surveillance Garbage Dump Site Surveillance Pollution Monitoring air quality monitoring and surveillance of pedestrian areas communicating pollution levels to technical staff / pedestrians Achim J. Lilienthal
1 Gas Distribution Modelling Applications Oil Refinery Surveillance Garbage Dump Site Surveillance Pollution Monitoring air quality monitoring and surveillance of pedestrian areas communicating pollution levels to technical staff / pedestrians Disaster Prevention Achim J. Lilienthal
1 Gas Distribution Modelling Applications Oil Refinery Surveillance Garbage Dump Site Surveillance Pollution Monitoring air quality monitoring and surveillance of pedestrian areas communicating pollution levels to technical staff / pedestrians Disaster Prevention Rescue Robots ... Achim J. Lilienthal
→ Contents Gas Distribution Mapping in Natural Environments – The Challenges Achim J. Lilienthal
2 Gas Distribution Mapping – Challenges Chaotic Gas Distribution diffusion advective transport turbulence video by Hiroshi Ishida Achim J. Lilienthal
2 Gas Distribution Mapping – Challenges Chaotic Gas Distribution Point Measurement sensitive sensor surface is typically small (often ≈ 1cm 2 ) Achim J. Lilienthal
2 Gas Distribution Mapping – Challenges Chaotic Gas Distribution Point Measurement Sensor Dynamics Achim J. Lilienthal
2 Gas Distribution Mapping – Challenges Chaotic Gas Distribution Point Measurement Sensor Dynamics Calibration complicated "sensor response ↔ concentration" relation dependent on other variables (temperature, humidity, ...) has to consider sensor dynamics variation between individual sensors long-term drift Achim J. Lilienthal
2 Gas Distribution Mapping – Challenges Chaotic Gas Distribution Point Measurement Sensor Dynamics Calibration Real-Time Gas Distribution Mapping changes at different time-scales rapid fluctuations slow changes of the overall structure of the average distribution Achim J. Lilienthal
→ Contents Kernel Based Gas Distribution Mapping Achim J. Lilienthal
3 Kernel Based Gas Distribution Mapping General Gas Distribution Mapping Problem x : 1 t given the robot trajectory : : 1 t 1 t ( | , ) p m x z gas gas Differences to Range Sensing calibration: readings do not correspond directly to concentration levels Achim J. Lilienthal
3 Kernel Based Gas Distribution Mapping General Gas Distribution Mapping Problem x : 1 t given the robot trajectory : : 1 t 1 t ( | , ) p m x z gas gas Differences to Range Sensing readings don't correspond directly to concentration levels chaotic gas distribution: an instantaneous snapshot of the gas distribution contains little information about the distribution at other times Achim J. Lilienthal
3 Kernel Based Gas Distribution Mapping General Gas Distribution Mapping Problem x : 1 t given the robot trajectory : : 1 t 1 t ( | , ) p m x z gas gas Differences to Range Sensing readings don't correspond directly to concentration levels instantaneous gas distribution snapshots contain little information about the distribution at other times point measurement: a single gas sensor measurement provides information about a very small area ( ≈ 1cm 2 ) Achim J. Lilienthal
3 Kernel Based Gas Distribution Mapping Time-Averaged Gas Distribution Mapping Problem x : 1 t given the robot trajectory : : av 1 t 1 t ( | , ) p m x z gas gas Kernel Based Gas Distribution Mapping interpret gas sensor measurements z t as random samples from a time-constant distribution assumes time-constant structure of the observed gas distribution randomness due to concentration fluctuations (measurement noise negligible) ⇒ kernel to model information content of single readings → Achim Lilienthal and Tom Duckett. " Building Gas Concentration Gridmaps with a Mobile Robot ". Robotics and Autonomous Systems, Vol. 48, No. 1, pp. 3-16, August 2004. Achim J. Lilienthal
3 Kernel Based Gas Distribution Mapping Analogy to Density Function Estimation estimate the PDF of a random variable (Parzen window) N 1 ( ) ∑ = − σ ˆ ( ) | |; p x K x x PW i Nh = 1 i K ← 2D univariate Gaussian from Duda, Hart, Stork " Pattern Classification " Achim J. Lilienthal
→ Contents Integrating SLAM and Gas Distribution Mapping Achim J. Lilienthal
4 Integrating SLAM and Gas Distribution Mapping General SLAM problem : : : 1 t t 1 t 1 t ( , | , ) p x m u z simultaneously estimate the map and the robot path given robot actions u and observations z Simultaneous Localisation and Gas Distribution / Occupancy Mapping ("GasSLAM") ( ) ← m = av , m m gas m occ ( ) ← z = , , z z z , t t gas t occ t Achim J. Lilienthal
4 Integrating SLAM and Gas Distribution Mapping The GasSLAM Problem general approach: inverse sensor model to estimate maps Achim J. Lilienthal
4 Integrating SLAM and Gas Distribution Mapping The GasSLAM Problem useful factorization if maps can be analytically estimated given a robot path hypothesis = ⇒ : : : 1 t t 1 t 1 t ( , | , ) p x m u z : : : : : : 1 t 1 t 1 t t 1 t 1 t 1 t ( | , ) ( | , , ) p x u z p m x u z estimate robot path compute maps using a particle filter analytically Rao-Blackwellization, Rao-Blackwellized Particle Filter (RBPF) Achim J. Lilienthal
4 Integrating SLAM and Gas Distribution Mapping GasSLAM – Map Computation observations z occ and z gas are conditionally independent assume independency between m occ and m gas Achim J. Lilienthal
4 Integrating SLAM and Gas Distribution Mapping GasSLAM – Map Computation observations z occ and z gas are conditionally independent assume independency between m occ and m gas ⇒ computing maps separately for each particle occupancy grid map using sensor integration [Moravec/Elfes 1985] gas distribution grid map using kernel based gas distribution mapping [Lilienthal/Duckett, 2004] determine max. likelihood estimate of the maps from the weighted average (using particle weights) Achim J. Lilienthal
4 Integrating SLAM and Gas Distribution Mapping GasSLAM – Estimation of the Robot Path sample from the motion model [i] [i] ~ ( | , ) x p x x u − t t t 1 t update weights with the observation model ω ∝ ω [i] [i] [i] [i] ( | , ) p z x m − 1 t t t t higher weights for particles that better corresponds with the current observations Achim J. Lilienthal
4 Integrating SLAM and Gas Distribution Mapping GasSLAM – Estimation of the Robot Path observation model = [i] [i] [i] [i] [i] ( | , ) ( , | , , ) p z x m p z z x m m , , t t gas t occ t t gas occ = [i] [i] [i] [i] ( | , ) ( | , ) p z x m p z x m , , gas t t gas occ t t occ ≈ η [i] [i] ( | , ) p z x m , occ t t occ occ occ use only the laser scanner gas to estimate the path Achim J. Lilienthal
→ Contents Experiments and Results Achim J. Lilienthal
5 Experiments Service Robot Sancho base: Pioneer 3DX laser range finder: SICK LMS 200 pair of e-noses Achim J. Lilienthal
5 Experiments Service Robot Sancho base: Pioneer 3DX laser range finder: SICK LMS 200 pair of e-noses Electronic Nose 4 metal-oxide gas sensors (Figaro): TGS 2600 [x2], TGS 2602, TGS 2620 sensors in a tube with CPU fan sampling frequency: 1.25 Hz separation: 14 cm; height: 11 cm Achim J. Lilienthal
Recommend
More recommend
