Bala Lakshminarayanan AUTOMATIC TARGET RECOGNITION April 1, 2004
Bala Lakshminarayanan Introduction Automatic Target Recognition ATR process –Detection –Tracking –Feature extraction –Identification / Recognition
Bala Lakshminarayanan Motivation Why ATR –Reduce human workload –Repetitive tasks –Limited vision of humans vs multi feature Where ATR
Bala Lakshminarayanan Objectives Aided ATR –Detect targets in high clutter environment –Low false alarm rate –High detection rate Autonomous ATR –High true positives –Ability to recognize target accurately –Consistency –LOAL, FAF
Bala Lakshminarayanan ATR…(1) Target Clutter Background variation, scene variation Target variations, new targets, Parameters Brightness, Temperature, Range/Distance, Velocity….
Bala Lakshminarayanan ATR…(2) Techniques involved –Sensor development –Algorithm development –Statistical pattern recognition –Adaptive learning –Neural networks –Image processing
Bala Lakshminarayanan ATR…(3) ATR classification By human-machine task sharing –Aided –Autonomous By range of output values –Binary –Multi valued
Bala Lakshminarayanan ATR…(3) Requirements –High resolution sensors –High speed processors –Collateral information –Low false positives –Real time operation –Recognition of new targets –Clutter independence
Bala Lakshminarayanan Sensors for ATR…(1) Visible camera – Brightness Infra red camera – Surface Temperature Acoustic – Distance RADAR – Range, velocity LASER – Range, 3D shape Microwave / Millimeter Wave – Range Multispectral Multi-sensor ATR
Bala Lakshminarayanan Sensors for ATR…(2) Active or passive sensors Criteria for sensors –All time operation –All weather operation –Range of sensor –Resolution –Parameter and ease of recognition
Bala Lakshminarayanan Sensors for ATR…(3) SensorTimeWeatherResolutionRangeParameter VisibleDay timeConstrainedLow/mediumLimitedBrightness FLIRDay/nightConstrainedLow/mediumMedium (10- 15km) Temperature AcousticDay/nightMedium dependent LowLimited (in meters) Distance LASERDay/nightConstrainedHighMedium (5km) Range/veloci ty/3D shape RADARDay/nightAllHigh Distance/vel ocity ….disadvantages of different sensors
Bala Lakshminarayanan Sensors for ATR…(4) New sensors –LADAR –SAR –Multi sensor
Bala Lakshminarayanan Problems in ATR Feature selection Algorithms for good recognition Measurement units for performance Computational power Representative databases –Orientation, time of day, weather, new targets, clutter, how much data, location, camouflage… Handling new targets (minimum distance classifiers) Overfitting
Bala Lakshminarayanan Performance measure…(1) Probability of detection Probability of classification (tracked/wheeled) Probability of recognition (tank/armored carrier) Probability of identification (brand name) False alarm rate
Bala Lakshminarayanan Performance measure…(2) SNR = (I t – I b )/I b –I t and I b are target and background intensities ROC –Plot of detection rate vs false alarm Confusion matrix Consistency
Bala Lakshminarayanan Performance measure…(3) Prob of detectionFalse alarm rate ATR SystemsMax Min Mean Human SystemsMax Min Mean
Bala Lakshminarayanan Performance measure…(4) Ground TruthSystemM60M113M35 M603/8 class0.67/ / /0.0.5 Human M1133/8 class0.08/ / /0.12 Human M353/8 class0.18/ / /0.36 Human Confusion matrix
Bala Lakshminarayanan Performance measure…(5) Improved measure Augustyn, “A new approach to Automatic Target Recognition” IEEE Trans on Aerospace and Electronic Systems
Bala Lakshminarayanan Learning in ATR…(1) ATR learning areas –Initial acquisition of domain theory –Adapt domain theory to new situations - “transfer” –Adapt new features Usually, supervised training occurs Need to use context based data
Bala Lakshminarayanan Learning in ATR…(2) Objectives of learning –Identify data to a class –Accommodate new features –Accommodate new targets –Express inability to classify (new target)
Bala Lakshminarayanan Learning in ATR…(3) ANNs –They model human brain –Feedforward or backpropagation networks –Backpropagation network is preferred since it is robust –Adaptive learning Limitations –Adapting to new situations is cumbersome –Highly sensitive to noise, occlusion –Nearest neighbor technique
Bala Lakshminarayanan Learning in ATR…(4) Explanation Based Learning –Machine derives explanation –4 inputs example, goal, operationality criterion (features), domain theory (relation) Examples can be generated using EBG –Irrelevant details removed from example –Explanation is generalized –Cannot learn new features –Difficult to implement
Bala Lakshminarayanan Learning in ATR…(5) Theory Revision –Refines domain knowledge –Knowledge engineer can provide approximate theory –Addressed deficiency of EBL
Bala Lakshminarayanan New developments…(1) Multi sensor ATR –Optical limits have been reached –Collateral information implies better results Data fusion –Information fusion –Pixel fusion –Decision fusion Model based systems
Bala Lakshminarayanan Questions & Comments