You are on page 1of 2

Particle Filters in the Localization of Mobile Robots

Submitted by Deepika Baskar for AE590 under Prof. Anouck Girard

I. Introduction
The problem localization involves estimating the relative pose of a robot in a given environment. It primarily equips
a robot with the capability of learning and translating in an environment autonomously. Particle filters are used to track
the quantities of interest using a non gaussian, sample based representation of the probability density function followed
by the association of weights to the particles of interest. Particle filters are recursive techniques of approximating the
posteriors over a discrete time. This report aims to provide an overview of the types of localization problems and the
implementation of particle filters in such scenarios.

II. Types of Localization problems


Depending upon the nature of the region where localization is performed, the problem of localization can be broadly
classified into four categories.
• Simple Position tracking - When the initial position of the robot is known and the localization is only limited to
estimating the drift in the odometry and making modifications to the final position estimates.
• Global Localization - This is a much more dramatic way of estimating the robot’s position in an environment
where the initial position of the robot is unknown.
• Kidnapped Robot - Similar to global localization but involves a well localized robot being introduced in an
unknown location in a manner that it assumes to be in its previously known location. It is usually regarded as one
of the hardest problems to solve.
• Multi- Robot Localization Problem - A swarm of robots attempt to localize as they interact among themselves to
reduce the dependency of the statistical estimates.

III. Implementation of Particle Filters


The particle filter estimates the position of the variable of interest in two phases : Predict and Update.
• Predict - To predict the probability distribution of the pose of a robot in motion, it important to incorporate even
the noise (usually modeled as Gaussian) that goes into the measurement of the position. Any motion can be
modeled as the sequence of Rotation-Translation-Rotation motion to a final pose given as follows:
 xt   xt−1   (∆s + ε2 ) cos (θ t−1 + α + ε1 ) 
     
 yt  =  yt−1  +  ∆s + ε2 ) sin (θ t−1 + α + ε1 )
     

 θ t   θ t−1 ∆θ + ε1 + ε3
     
  
     
where εx represents the gaussian measurement noise.
• Resampling - Over the course of actions the weights on the particles that do not contribute the final position
estimate have dropped considerably and thus leading to the reduction of number of particles. In order to improve the
reliability in the filtering process it is essential that optimal number of particles are sampled for pdf representation.
• Update - After the motion of the robot the tracker present in the robot updates the pose of the robot. This paired
up with a sensor model update the log odds of the occupancy grid to provide with the position estimate of the map
that has evolved over time. For the stationary position xs of the robot, if the measurements from the sensors are
give by z = (ρ, θ̂, φ̂) as follwing, p
 ρ   dx 2 + dy 2 
   
 θ̂  =  atan 2(dy, dx) − θ̂ m 
   
 φ̂   atan 2(−dy, −dx) − θ̂ s 
   
   

1
Fig. 1 The figures (a), (b) and (c) depict the evolution of particles over time using MCL technique.

then the estimated robot pose is give by,


 
 xm   xs + ρ ∗ cos φ̂ + θ̂ s 
ext
   
 
 =  ys + ρ ∗ sin φ̂ + θ̂ s

 yme x t
   

 θ̂ me s t
   
π + φ̂ + θ̂ s − θ̂
  
   
Monte Carlo Localization has widely been studied and used in the field of mobile robots localization. It is found to
be efficient with an optimal resolution of grid and size of the grid cells. This is because smaller sample size results in
loss in accuracy and larger sizes lead to longer computation time.

IV. Conclusion
This brief report summarizes the general steps involved in localization of a mobile robot using statistical inference
approach. Additionally, the challenge of choosing an optimal sample size was also analyzed and suitable arguments
were presented.

You might also like