The presentation with the topic AI methods for localization in a noisy environment, held by Ana Antonova and Kameliya Kosekova, was introduced at Robotics Days '19.
In the next slides, you can find information techniques for Robot localization in more details and several GitHub Repos on the topic.
3. LOCALIZATION?
Position Tracking Global
Localization Kidnapped Robot
Static and Dynamic Environment
Passive and Active Localization
Localization and Mapping
A.ANTONOVA & K.KOSEKOVA 3
8. GAUSSIAN FILTERS
Earliest implementations of the Bayes filter
Assumptions:
Beliefs are represented by multivariate normal distributions
Properties of Gaussians:
Unimodal
A.ANTONOVA & K.KOSEKOVA 8
9. KALMAN FILTER
Applicable to linear systems and continuous states
Assumptions:
Linearity in the state probability
Linearity in the measurement probability
Normally distributed initial belief
A.ANTONOVA & K.KOSEKOVA 9
10. EXTENDED KALMAN FILTER
Applicable to nonlinear systems
Applies linearization via Taylor expansion
Approximates the nonlinear functions -> leads to a
Gaussian posterior belief
The most popular tool for state estimation
Computationally efficient
Drawbacks: Uncapable of representing multi-modal
beliefs
A.ANTONOVA & K.KOSEKOVA 10GitHub: AtsushiSakai/PythonRobotics
11. EKF LOCALIZATION
Special case of Markov localization
Well-suited technique for local position tracking with limited uncertainty and in environments with distinct
features (landmarks)
Initial position is known
For the implementation of the algorithm we need the following:
Motion model
Measurement model
Map of the environment
A.ANTONOVA & K.KOSEKOVA 11
12. PARTICLE FILTER
Approximates the posterior by a finite number of
parameters
A random state samples (particles) drawn from the
posterior
Each particle is a hypothesis to the true state
Uses importance resampling to form a set of particles
This set of particles approximates the belief
A.ANTONOVA & K.KOSEKOVA 12
Jeremy Cohen, “Self-Driving Cars & Localization”
13. MONTE CARLO LOCALIZATION
Uses particle filter
Able to solve local/global localization and kidnapped robot problems (recover from localization failure)
The kidnapped robot problem is solved by adding random particles
Able to process raw sensor measurements and negative information
The accuracy-computational costs trade-off is achieved through the size of the particle set
For the implementation of the algorithm we need the following:
Motion model
Measurement model
Map of the environment
Initial belief
A.ANTONOVA & K.KOSEKOVA 13
14. SLAM
Robot has no initial information of the environment
Online and full SLAM
Known correspondence and unknown
correspondence
A.ANTONOVA & K.KOSEKOVA 14
15. EKF SLAM
Underwater vehicle Oberon, developed at the University of Sydney. Image
courtesy of Stefan Williams and Hugh Durrant-Whyte.
Earliest application of SLAM
Some assumptions:
Feature-based maps
Gaussian noise
Positive measurements
Online SLAM
A.ANTONOVA & K.KOSEKOVA 15
16. FASTSLAM
Resolving the computational complexity of EKF SLAM
Each Particle in the algorithm get separate landmark
estimators for each landmark – log(N) time.
A.ANTONOVA & K.KOSEKOVA 16
GitHub: AtsushiSakai/PythonRobotics
17. 17
“FastSLAM: A Factored Solution to theSimultaneous Localization and Mapping
ProblemWith Unknown Data Association”, M. Montemerlo