# Learning Algorithms for Terrain Analysis

Document Sample

```					Learning Algorithms for
Terrain Analysis
Jackie Soenneker
The Scenario

 Robot uses its LADAR to scan an area
 “Features” are extracted from the scan data
 Features are passed to a terrain classifier
 Terrain classifier returns the type of terrain in
the scan
The Terrain Classifier

 My part: the terrain classifier
 The terrain classifier will use a learning
algorithm
 Recommended learning algorithm: a Neural
Net
Learning Methods

 Neural Networks
Consist of an input layer, output layer, and hidden
layer(s)
Input values feed through the network, influenced
by connection weights
Network learns by adjusting the connection weights
Learning Methods con.

 Maximum Likelihood Classifier
An input is labeled as an element of the class that it
most likely belongs to
Class likelihoods are based upon a Probability
Density Function (PDF; represents the probability
distribution), which is learned from training data
 Bayesian Classifier
Basically the same as the ML classifier
Research Sample
Paper              Learning Method(s)           Task              Data Source

1*                 Back-prop NN      Terrain classification     SAR (radar)

2                Max. Likelihood    Terrain classification    Color images

3                BP NN and ML       Terrain classification       Landsat

4             Bayesian classifier   Terrain cover class.         LADAR

5             Dynamic Learning      Terrain cover class.           SPOT
NN                                       (visible/near-IR)
6                 Back-prop NN      Terrain classification    Color images

7                 Back-prop NN      Obstacle detection           LADAR

* See last slide for paper titles, etc.
Justification for Using NNs

 Both ML and Bayesian classifiers need a
PDF; NNs do not
 Several papers compared performance of
NNs and ML classifiers; NNs usually did
better
 NNs are good at dealing with noisy data and
generalizing to new situations
NN for Our Project
Inputs        Hidden Layer(s)         Outputs

height

slope                                          grass

mean
concrete

standard
deviation
trees

distribution
coefficient

weighted connections
Conclusions

 Nobody else is doing exactly what we are
 Neural Nets and Maximum Likelihood and
Bayesian classifiers are most popular
learning methods for terrain analysis
 Neural Nets have characteristics that may be
 I recommend that we use a Neural Net!
Table Paper Key

   1 – Decatur, Scott Evan, “Application of Neural Networks to Terrain Classification,” in Proceedings
of the IEEE International Joint Conference on Neural Networks, pp. 283-288, June 1989.
   2 – P. Jansen, W. van der Mark, J.C. van den Heuvel, and F.C.A. Groen, “Colour based Off-Road
Environment and Terrain Type Classification,” in Proceedings of the 8th International IEEE
Conference on Intelligent Transportation Systems, 2005.
   3 – H. Bischof, W. Schneider, and A.J. Pinz, “Multispectral Classification of Landsat-Images Using
Neural Networks,” in IEEE Transactions on Geoscience and Remote Sensing, 1992.
   4 – N. Vandapel, D.F. Huber, A. Kapuria and M. Herbert, “Natural Terrain Classification using 3-D
Ladar Data,” in Proceedings of the 2004 IEEE International Conference on Robotics and
Automation, 2004.
   5 – K.S. Chen, S.K. Yen and D.W. Tsay, “Neural classification of SPOT imagery through
integration of intensity and fractal information,” International Journal of Remote Sensing, 18:4,
763-783, 1997.
   6 – I. Davis, Neural Networks for Real-Time Terrain Typing, tech. report CMU-RI-TR-95-06,
Robotics Institute, Carnegie Mellon University, January, 1995.
   7 – I.L. Davis and A. Stentz, “Sensor Fusion for Autonomous Outdoor Navigation Using Neural
Networks,” tech. report CMU-RI-TR-95-05, Robotics Institute, Carnegie Mellon University, January
1995.

```
DOCUMENT INFO
Shared By:
Categories:
Stats:
 views: 7 posted: 6/1/2010 language: English pages: 10
How are you planning on using Docstoc?