Engler and Prantl system of classification in plant taxonomy
IROS_2019_Naoki_Akai
1. Misalignment Recognition Using Markov
Random Fields with Fully Connected Latent
Variables for Detecting Localization Failures
Nagoya University, Japan
Naoki Akai, Luis Yoichi Morales, Takatsugu Hirayama, and Hiroshi Murase
Accepted by The IEEE Robotics and Automation Letters
2. Background
⚫ Accurate localization is fundamental for precise autonomous navigation
• Autonomous navigation systems are usually implemented with an
assumption that accurate localization is always available, e.g., [1]
• Autonomous navigation will be failed once the localization has failed
[1] N. Akai et al., “Autonomous driving based on accurate localization using multilayer LiDAR and dead reckoning,” In Proc. of the IEEE ITSC, 2017.
3. Motivation
⚫ Guarantee safety of localization-based autonomous navigation systems
• Immediate and exact localization failure detection
• Estimate reliability of localization results
⚫ Target
• The 2D LiDAR-based localization problem
• Without a redundant poisoning system, i.e., other sensors like
GPS and camera are not used
4. Difficulty of the failure detection
⚫ The independent measurement assumption (IMA)
• The IMA is needed to practically solve the localization problem
• The IMA enables to decompose the measurement model as:
• E.g., Y is point cloud and yk is a point (x and m are pose and map)
⚫ The entire relation of the measurements is ignored
5. Difficulty of the failure detection
⚫ Cannot intuitively detect localization failures because of the IMA
• We humans can intuitively recognize the failure (bottom left)
• Detecting the failure using the measurement model, e.g. beam
model [2], is not intuitive (bottom right)
[2] S. Thrun et al., “Probabilistic robotics,” MIT Press, 2005.
6. Proposed method
⚫ Introduce to a new framework which can consider the entire relation
• Use Markov random fields with fully connected latent variables
Consider the entire relation
via the full connected latent
variables
Residual errors (observable)
7. ⚫ Assume a residual error is generated from a probabilistic distribution
• The residual errors of success and failure cases are according to the
Gaussian and exponential distributions, respectively
• The latent variables indicate the probabilistic distributions
• Estimate the posterior distribution over the latent variables
Success case Failure case
Proposed method
8. ⚫ Define the failure probability using the posterior distribution
• The k-the latent variable is denoted as:
• If , the k-th measurement is misaligned
• If , the k-th measurement is obtained from unknown objects
Misalignment ratio and its threshold
Proposed method
aligned
9. Simply say the proposed method
⚫ The proposed method distinguishes weather the sensor measurements
are aligned or misaligned to a map with the probabilistic manner
• The failure is detected based on the misalignment estimation
• The full connections improve the misalignment estimation accuracy
Failure probability
Estimation errors
Success localization sample
Unknown
obstacles
Aligned
measurements
5 m
Unknown
obstacles
Failure probability
Estimation errors
Failure localization sample Misaligned
measurements
5 m
11. ⚫ The localization failures can be immediately detected
Simulation demonstrations
12. ⚫ The large estimation errors can also be detected
Simulation demonstrations
13. Demonstrations with the real robot
⚫ The proposed method also works in the real environment
Hokuyo 2D LiDAR
(UTM-30LX)
14. Dataset example for comparison experiments
⚫ Datasets composed of success and failure localization samples are
created using the 2D localization simulator in five indoor environments
Success sample Failure sample
Slight
mismatches
15. Comparison methods
⚫ Two model- and four leaning-based methods are compared
• Model-based method
➢ Markov random fields w/o full connection, Root mean squared
• Machine learning-based methods
➢ SVM, Logistic regression, AdaBoost, CNN [3]
[3] S. Zagoruyko et al., “Learning to compare image patches via convolutional neural networks,” In Proc. of the IEEE CVPR, 2015.
Residual
errors
Regress the
failure probability
CNN architecture
16. Comparison results
⚫ Performances of the proposed method and CNN are close
• The accuracy by the proposed methods exceeded 95 %
• Totally, the proposed method outperformed the all of the method
17. Comparison of w and w/o the full connections
⚫ The full connections improve the misalignment recognition accuracy
• Compared them while changing the misalignment ratio threshold
• The proposed method was not affected by the change of the
threshold because the misalignments were exactly recognized
• Performance of the method w/o the full connections significantly
depends on the threshold
18. Conclusion
⚫ Proposed the use of the Markov random fields with fully connected
latent variables for detecting localization failures
• The entire measurement relation is ignored to solve the localization
problem due to the use of the independent measurement assumption
• The full connection is able to consider the entire measurement
detection for estimating the misalignment measurements
• Based on the misalignment recognition, the localization failures can
be immediately detected
• The failure detection accuracy by the proposed method exceeded
95 % and outperformed the compared methods