Download presentation
Presentation is loading. Please wait.
Published byAmbrose Brooks Modified over 9 years ago
1
Behavior Recognition of Autonomous Underwater Vehicles For CS 7631: Multi Robot Systems Michael “Misha” Novitzky School of Interactive Computing Georgia Tech
2
Motivation
3
Yellowfin
4
Yellowfin Untethered
5
The Papers Behaviour recognition for spatially unconstrained unmanned vehicles, R. Baxter, D. Lane, and Y. Petillot, IJCAI 2009 Conditional Random Fields for Behavior Recognition of Autonomous Underwater Vehicles, M. Novitzky, C. Pippin, T. Collins, T. Balch, and E. West, under review at IROS, 2012.
6
Baxter et al. : Problem Description How do we determine the behavior of an Unmanned Underwater Vehicle with only GPS coordinates?
7
Baxter et al.: Insights Break from location dependent GPS coordinates and use compass heading Behaviors encoded as: W – W – NW – N – NE -E
8
Baxter et al.: Approach Took self-localization data from post UAV missions Converted GPS coordinates to compass heading
9
Baxter et al.: Approach X1 y1 X2 y2 y3 Hidden Markov Models (HMMs) Given: Example Sequence Learn: Transition Probabilities Emission Probabilities
10
Baxter et al.: Approach Run 8 HMMs (one for each behavior) HMM with highest negative log-likelihood for K consecutive time slices = WINNER!
11
Baxter et al.: Implementation All trajectories were created via simulation Obviously, not using a real robot implies that this may not work in situ
12
Baxter et al.: Experiments Simulated UAV trajectories Added noise on encoding such as: N -> NW
13
Baxter et al.: Results Precision
14
Baxter et al.: Results Confusion under 70% noise
15
Baxter et al.: Critique Authors demonstrated a system for behavior classification Variable length testing Implementation – restricted vehicles to compass directions – which is not really location agnostic Only in simulation – will this work on real UUV?
16
Novitzky et al.: Problem Description Can we recognize the behaviors of UUVs using two different approaches? Which is best? When?
17
Novitzky et al.: Insights Using environmentally agnostic encoding method Use real sonar data CRFs vs HMMs more accurate?
18
Novitzky et al.: Approach Environment agnostic discretization:
19
Novitzky et al.: Approach HMMs: one HMM per behavior The largest negative log-likelihood is the WINNER! X1 y1 X2 y2 y3
20
Novitzky et al.: Approach X1 X2 Y Y Conditional Random Fields (CRFs) Given: Example Sequences Learn: Potential Functions One CRF: Each X is a label Y’s include all instances
21
Novitzky et al.: Implementation Simulation Real sonar data YellowRay ROV All analyzed using MATLAB
22
Novitzky et al: Experiments Stationary Observer: Simulation 600 Train 400 Test Real Sonar Data
23
Novitzky et al: Experiments Simulated: Track & Trail
24
Novitzky: Results
25
Novitzky et al.: Critique Not variable length testing Not enough real sonar data Simulated noise accurate? Guassian? Actually use real vehicles and data!
26
Comparison of the Two Encoding methods Baxter et al. variable length testing Baxter et al. has more behaviors Novitzky et al. has real sonar data if have ample training data use CRFs else use HMMs
27
Questions? Paul Robinette Andrew Melim
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.