170 likes | 226 Views
Explore the use of Bayes filters like Kalman filter and Particle filter for location estimation in pervasive computing scenarios. Learn about implementation examples and experiment results with sensors and tracking techniques.
E N D
Bayesian Filtering for Location Estimation Authors: Dieter Fox, Jeffrey Hightower, Lin Liao, Dirk Schulz, Gaetano Borriello -- PERVASIVE computing 2003
Outline • Motivation • Bayes filters • Implement Bayes filters • Kalman filter • Multi-hypothesis tracking • Grid-based approaches • Topological approaches • Particle filter • Experiment
Motivation • No location sensor takes perfect measurements or works well in all situations • Representing and operating on uncertainty with a statistical tool will benefit the measurements • Estimating location information is the most fundamental in many pervasive computing scenarios • Representing locations statistically enables a unified interface for location information, independent of the sensors used
Bayes filters • Bel(xt) = P(xt | z1, z2,…, zt) = α P(zt | xt) * ∫ P(xt | xt-1)* Bel(xt-1)dxt-1 = f(zt, Bel(xt-1)) • Bel(xt) : state distribution at time t • α : a normalizing constant • P(zt | xt) : measurement model • P(xt | xt-1) : state transition model
Bayes filters • P(xt | z1:t) = P(z1:t-1, zt | xt) * P(xt) / P(z1:t) = P(zt | xt, z1:t-1) * P(z1:t-1 | xt) * P(xt) / P(z1:t) = P(zt | xt) * P(z1:t-1 | xt) * P(xt) / P(z1:t) = P(zt | xt) * P(xt | z1:t-1) * P(z1:t-1) / P(z1:t) = P(zt | xt) * P(xt | z1:t-1) * P(z1:t-1) / (P(zt | z1:t-1) * P(z1:t-1)) = P(zt | xt) * P(xt | z1:t-1) / (P(zt | z1:t-1) = α * P(zt | xt) * P(xt | z1:t-1) = α * P(zt | xt) * ∫P(xt | xt-1) * P(xt-1 | z1:t-1) dxt-1
Implement Bayes filters • Require specifying • the measurement model P(zt | xt), • the state transition model P(xt | xt-1) , • and the representation of the belief Bel(xt) • Implementation examples • Kalman filters • Multi-hypothesis tracking • Grid-based approaches • Topological approaches • Particle filters
Kalman filters • Represents the belief as Gaussian distribution • Bel(xt) is Gaussian • Measurement model is linear function • State transition model is linear function • Advantage : • Computational efficiency, using efficient matrix operations on the mean and covariance • Disadvantage : • Representational power, can represent only Gaussian distribution
Multi-hypothesis tracking • Represents the belief as mixtures of Gaussian • wt(i) is proportional to the sensor measurements • Each hypothesis using a Kalman filter • Advantage : • more widely applicable than the Kalman filter • Disadvantage : • computationally more expensive • Require sophisticated techniques or heuristics to determine when to add or delete hypotheses
Grid-based approaches • Represents the belief on discrete, the integration in equations will replace to summation • For indoor location estimation, grid-based filters tessellate the environment into small patches • Advantage : • Can represent distributions over the discrete state space • Disadvantage : • Computational and space complexity are high
Topological approaches • Using a graph to represent the environment, each node is a location and the edges is the environment’s connectivity • Advantage : • Efficiency, because they represent distributions over small, discrete state spaces • Disadvantage : • The representation is coarseness • Require sensors related to the environment’s layout
Particle filters • Represent belief by sets of particles • xt(i) is a state, wt(i) is weight • Resampling by state transition model • Weighted by measurement model • Advantage : • Can represents probability densities • Can work on non-Gaussian, non-linear dynamic systems • Very efficient than grid-based approach, focus only on state space with high probability • Low implementation overhead • Disadvantage : • Worst case complexity grows exponentially, must careful when applying to high dimension estimation problems
Experiment • Sensors • Ultrasound sensors & tags • 4.5-meter, Gaussian distribution of measured distance • Infrared sensors & badges • A no distance detection, in a specific area • Laser range finders • Several short beams form a shadow region indicating a person’s presence
Experiment • Implement approach • Particle filters • The state contains person’s location, orientation, and motion velocity
Experiment • Constrain the state space to locations on a Voronoi graph, which is a structure similar to a skeleton of an environment’s free space
Expriment • Tracking multiple people • Problem • Requires maintaining the hypotheses for possible track continuations • Proposed solution • Track individual people using Kalman filters • A particle filter maintains multi hypotheses regarding the ID of people