Title: Clustering%20Sensor%20Data%20for%20Terrain%20Identification%20using%20a%20Windowless%20Algorithm
1Clustering Sensor Data for Terrain Identification
using a Windowless Algorithm
- Philippe Giguère
- Gregory Dudek
2Outline
- Problem
- Proposed algorithm
- Evaluation on synthetic datasets
- Application to terrain identification
- Conclusion, future work
3Exploring the World
Richard Altendorfer, Ned Moore, Haldun Komsuoglu,
Martin Buehler, H. Benjamin Brown Jr., Dave
McMordie, Uluc Saranli, Robert Full, Daniel E.
Koditschek RHex A Biologically Inspired Hexapod
Runner. Auton. Robots (AROBOTS) (2001)
4Autonomous Terrain Identification
5UnsupervisedLearning of Terrains
6UnsupervisedLearning of Terrains
7Assumptions
- Representative features
- Time continuity
8Time Series Clustering
- Time-Window
- J. Kohlmorgen and S. Lemm. An on-line method for
segmentation and identification of non-stationary
time series. In IEEE Neural Networks for Signal
Processing Workshop 2001. - S. Lenser and M. Veloso. Automatic detection and
response to environmental change. In Proceedings
of the ICRA, May 2003. - Regularization (smoothness)
- J. Kohlmorgen, S. Lemm, G. Rätsch, and K.-R.
Müller. Analysis of non-stationary time series by
mixtures of self-organizing predictors. In IEEE
Neural Networks for Signal Processing Workshop,
2000.
9Potential Issues
- Time-windows
- Transitions
- Complex-shaped distributions
- Regularization
- Right regularization constant l?
10Classifier Definition
11Classifier Training
12Classifier Training
13Proposed Algorithm
- A classifier generalizing
over local distances in feature space - e.g., k-Nearest Neighbors
- Training rule
14Optimization Simulated Annealing
Set q rand while (temp gt finaltemp) qnew
RandomChange(q) if E(qnew,X) lt E(q,X) q
qnew Always accept good elseif
P(E(q,X),E(qnew,X),temp) gt rand q qnew
Sometimes accept bad end decrease temp
Cooling end return q
15Testing on Synthetic Datasets
- Classifiers
- Linear Separator
- Mixture of Gaussians
- K-Nearest Neighbors
- Distributions
16Synthetic Datasets Generation
17Synthetic Datasets Generation
18Linear Separator
Segment Length 3
19Synthetic Datasets
Number of classes in (bracket)
20k-Nearest Neighbors
k 30
Symbol Cluster Assignment
Segment Length 5
21Played-back through aSliding Window
22Comparison withWindow-based Clustering
Results
S. Lenser and M. Veloso. Automatic detection and
response to environmental change. In Proceedings
of the ICRA, May 2003.
23AQUA Robot (Based on RHex)
G. Dudek, M. Jenkin, C. Prahacs, A. Hogue, J.
Sattar, P. Giguere, A. German, H. Liu, S.
Saunderson, A. Ripsman, S. Simhon, L. A.
Torres-Mendez, E. Milios, P. Zhang, I. Rekleitis.
A Visually Guided Swimming Robot, Proceedings of
the 2005 IEEE/RSJ International Conference on
Intelligent Robots and Systems, pp. 1749-1754,
2005.
24AQUA Sensors
25Features Dimensionality Reduction
26Features Dimensionality Reduction
27Features Dimensionality Reduction
28Features Dimensionality Reduction
29Terrains
- Robot manually driven over
- Grass
- Tilled Earth
- Snow
- Ice
- Linoleum
302 Terrains (Linear Separator)
315 Terrains Data
325 Terrains Data
335 Terrains (Gaussian Mix. Model)
Classification Rate 91
345 Terrains (Gaussian Mix. Model)
35Conclusion
- Clustering combining time and feature
- No
- Time window
- Regularization
- Training a classifier with cost function
- Unsupervised learning much easier
- Experimentally validated on hexapod robot
36Future Work
- More terrains
- Accelerated optimization
- On-line version
- Finding number of clusters
- Other problems
- Texture segmentation
37Questions?
38Comparing with HMM
39Dimensionality Reduction
401st Part Finding Projection
41Cost Function Formulation
42k-NN Few Points
Segment Length 3
43k-NN 3-to-1 Ratio
Segment Length 3
44k-NN Overlapping
Segment Length 3
45k-NN 6 Classes
Segment Length 3
46Clustering Sensor Data for Terrain Identification
using a Windowless Algorithm
- Philippe Giguère
- Gregory Dudek
47Outline
- Problem
- Proposed algorithm
- Evaluation on synthetic datasets
- Application to terrain identification
- Conclusion, future work
48Exploring the World
Richard Altendorfer, Ned Moore, Haldun Komsuoglu,
Martin Buehler, H. Benjamin Brown Jr., Dave
McMordie, Uluc Saranli, Robert Full, Daniel E.
Koditschek RHex A Biologically Inspired Hexapod
Runner. Auton. Robots (AROBOTS) (2001)
49Autonomous Terrain Identification
50UnsupervisedLearning of Terrains
51UnsupervisedLearning of Terrains
52Assumptions
- Representative features
- Time continuity
53Time Series Clustering
- Time-Window
- J. Kohlmorgen and S. Lemm. An on-line method for
segmentation and identification of non-stationary
time series. In IEEE Neural Networks for Signal
Processing Workshop 2001. - S. Lenser and M. Veloso. Automatic detection and
response to environmental change. In Proceedings
of the ICRA, May 2003. - Regularization (smoothness)
- J. Kohlmorgen, S. Lemm, G. Rätsch, and K.-R.
Müller. Analysis of non-stationary time series by
mixtures of self-organizing predictors. In IEEE
Neural Networks for Signal Processing Workshop,
2000.
54Potential Issues
- Time-windows
- Transitions
- Complex-shaped distributions
- Regularization
- Right regularization constant l?
55Classifier Definition
56Classifier Training
57Classifier Training
58Proposed Algorithm
59Optimization Simulated Annealing
60Testing on Synthetic Datasets
- Classifiers
- Linear Separator
- Mixture of Gaussians
- K-Nearest Neighbors
- Distributions
61Synthetic Datasets Generation
62Synthetic Datasets Generation
63Linear Separator
Segment Length 3
64Synthetic Datasets
Number of classes in (bracket)
65k-Nearest Neighbors
k 30
Segment Length 5
Symbol Cluster Assignment
66Played-back through aSliding Window
67Comparison withWindow-based Clustering
Results
S. Lenser and M. Veloso. Automatic detection and
response to environmental change. In Proceedings
of the ICRA, May 2003.
68AQUA Robot (Based on RHex)
G. Dudek, M. Jenkin, C. Prahacs, A. Hogue, J.
Sattar, P. Giguere, A. German, H. Liu, S.
Saunderson, A. Ripsman, S. Simhon, L. A.
Torres-Mendez, E. Milios, P. Zhang, I. Rekleitis.
A Visually Guided Swimming Robot, Proceedings of
the 2005 IEEE/RSJ International Conference on
Intelligent Robots and Systems, pp. 1749-1754,
2005.
69AQUA Sensors
70Features Dimensionality Reduction
71Features Dimensionality Reduction
72Features Dimensionality Reduction
73Features Dimensionality Reduction
74Terrains
- Robot manually driven over
- Grass
- Tilled Earth
- Snow
- Ice
- Linoleum
752 Terrains (Linear Separator)
765 Terrains Data
775 Terrains Data
785 Terrains (Gaussian Mix. Model)
Classification Rate 91
795 Terrains (Gaussian Mix. Model)
80Conclusion
- Clustering combining time and feature
- No
- Time window
- Regularization
- Training a classifier with cost function
- Unsupervised learning much easier
- Experimentally validated on hexapod robot
81Future Work
- More terrains
- Accelerated optimization
- On-line version
- Finding number of clusters
- Other problems
- Texture segmentation
82Questions?