a uav vision system for airborne surveillance
Post on 19-Jan-2016
64 Views
Preview:
DESCRIPTION
TRANSCRIPT
A UAV Vision System for Airborne Surveillance
M.Kontitsis, K. Valavanis
IEEE 2004 International Conference on Robotics and Automation
Technical University of Crete
University of South Florida
N. Tsourveloudis
Objectives
• Present a methodology for the design of a machine vision system for aerial surveillance by Unmanned Aerial Vehicles (UAVs)
• Identify specified thermal source• Perform these functions on board the
UAV in Real time• Flexible enough to be used in a variety
of applications
Machine Vision System
IR/NIR image Noise reduction
Feature extraction
(Size, Mean intensity)
Feature vectors classification
Alarm on/off
Persistence
Input Images
IR (3μm ~ 14μm)
8bit grayscale
Near IR camera (1μm ~ 3μm)
Noise Reduction
5x5 spatial Gaussian filter
+ Smoothes noise while preserving most of the features on the image
Feature Extraction
Size of region using a region growing algorithm
Mean intensity of region defined as
This module attempts to extract information about the regions on the image
Feature Vector Classification Subsystem
Mean Intensity of Region
Size of Region
Target Identification Possibility
Fuzzy Classifier
Mean Intensity Membership Functions
Grayscale values
HighMidLow
Region Size Membership Functions
Pixels
Small Medium Large
Objective ID Possibility Membership Functions
Output of the Fuzzy Classifier
Classification Example
Classification Result
p>0.8
0.5<p<0.8
p<0.5
Alarm raising
Persistent classification of a certain region as of High Possibility raises the alarm
The region that raised the alarm is pin-pointed by a red cross
The alarm stays on even if the thermal source is temporarily occluded by surroundings or lost due to violent camera vibrations
Alarm raising
Mechanism used : Alarm Registry
Region Coordinates Variable persistancei1,j1 p1
i2,j2 p2
…. ……..
in,jn pn
If pi > Ton => Activate alarm
If pi < Toff => Deactivate alarm
Complexity
Noise Reduction O(n2) for (nxn) image
Region Growing O(n2) for (nxn) image
Fuzzy Logic Classifier* O(nxm)
*in its current implementation
n inputs, m rules
Case Study: Forest fires
Adjusting membership functions manually
Classification Example (objective present)
Thermal source (fire)
Classification Result (objective present)
possibility>0.7 0.5< possibility <0.7 possibility <0.5
Classification Example (objective absent)
Classification Result (objective absent)
possibility>0.7 0.5< possibility <0.7 possibility <0.5
Classification Result (Video) (objective present)
Classification Result (Video) (objective absent)
Automatic Parameter Selection
aij bij cij dij for i=1 and j =1,2,3 which define the form of the membership functions of Mean Intensity
aijbij cij dij
Basic Elements of the Genetic Algorithm
Fitness function
fitness(x)=1 correct deactivation of the alarm
fitness(x)=0 in any other case
correct activation of the alarm
Chromosome => parameters x=(aij bij cij dij)
Basic Elements of the Genetic Algorithm
Selection operator selects individuals for mating as many times as the ratio of their fitness to the total fitness of the population
Crossover operator crossover probability pc=0.7
Mutation operator mutation probability pm=0.001
Mean Intensity M. F. as evolved by GA
Result (using GA for parameter selection)
Remarks
Adjustable for a variety of applications
Real time execution
Correct identification rate of about 90%
False alarms not entirely avoided (especially in the system evolved by the GA)
top related