AUTOMATIC TARGET RECOGNITION OF CIVILIAN TARGETS September 28 th, 2004 Bala Lakshminarayanan
Outline Objective Introduction to ATR Details of SFTB Database creation Segmentation Feature extraction, classification Results Conclusions
Objective Civilian target classification Sensor fusion SFTB objectives - Generation of dataset for ATR - Ground truth data collection
Introduction to ATR What is ATR Why do we need it Types of ATR - Aided, unaided - Binary, multi-valued Problems
Introduction to ATR Requirements - Real time operation - Low false positives - High detection rates Applications - Military - Medical - Industrial
SFTB Nodes - Base station - 2 with IR sensor - 1 with visible light sensor Node placement Targets (cars, light trucks, SUVs) Ground truth collection equipment Scenarios
SFTB Image provided by Night vision lab
SFTB Fully exposed targets except by other presence on scene Stationary sensors Daylight operation License plates not readable Constant velocity/acceleration Different scenarios (3) Simultaneous data capture
Images Node 1 Node 3 Node 2
Project Objective Use IR and visual images to classify targets Use sensor fusion to improve accuracy Creation of image database Creation of framework Segmentation, feature extraction, classification
Database Creation Images in.arf files Use frames captured at same time “Event start” - Range from Node2 = 20 “Event end” - Outside FoV of Node3
Framework Start Grab frame from dataset filename() Segment bgSubtract(), motionDet() Extract features invMoment() Classify readData(), knn() End Inputs-nodeID, scenario…
Segmentation Used to identify the target/RoI in the frame Methods - Thresholding - Background subtraction - Motion based segmentation
Segmentation Background subtraction median(frame)-median(background) Noise removal by neighbourhood() -= - =
Segmentation Motion based segmentation temp1=average(prev)-average(frame) temp2=average(next)-average(frame) temp1&temp2
Feature Extraction Features should describe similar targets similarly Seven invariant moments (Hu, 1962) Computed from central moments, third order Translational invariance – C.G Distance invariance – Size normalization
Feature Extraction 1 = 20 + 02, 2 = ( 20 - 02 ) 2 11 3 = ( 12 ) 2 + ( 21 ) 2, 4 = ( 30 + 12 ) 2 + ( 03 + 21 ) 2 5 = (3 12 )( 30 + 12 )[( 30 + 12 ) 2 –3( 21 + 03 ) 2 ] + (3 21 - 03 )( 21 + 03 ) [3( 30 + 12 ) 2 – ( 21 + 03 ) 2 ] 6 = ( 20 - 02 )[( 30 + 12 ) 2 – ( 21 + 03 ) 2 ] + 4 11 ( 30 + 12 )( 21 + 03 ) 7 = (3 21 - 03 )( 30 + 12 )[( 30 + 12 ) 2 - 3( 21 + 03 ) 2 ] + (3 12 - 30 )( 21 + 03 ) [3( 30 + 12 ) 2 – ( 21 + 30 ) 2 ] Central moments Normalized moments
Classification Supervised or unsupervised k-nearest neighbour method Training vectors are given Find k nearest neighbours, maximum presence
Results 3 classes - 1, 2, 4; single scenario 7 features - 5 training vectors, 2 testing vectors k = 1, 3 Class K= K= Class K= K=
Results Overall classification results k=1 – 58.33% k=3 – 50% Target1 – 25% Target2 – 38.5% Target4 – 100%
Results Confusion matrix k= k=
Conclusions Database created Basic framework has been laid Robust segmentation needed More training vectors Segmentation does not work for px files
Future work Segmentation - Quadtree based split-merge - Use of Kalman filters - Histogram based segmentation Better features need to be used
Thanks ?? and !!