3-D OBJECT RECOGNITION IN A NEURAL NETWORK THAT PERFORMS SENSOR FUSION

THE FUSION OF SENSOR DATA TO PERFORM 3-D TARGET RECOGNITION FROM 2-D VIEWS IS AN INHERENTLY COMPLEX, COMPUTE-BOUND PROBLEM CHARACTERIZED BY AN OVERWHELMING AMOUNT OF PIXEL DATA WHICH MUST BE PROCESSED AND INTEGRATED INTO A SINGLE COHERENT DESCRIPTION OF OBJECTS IN AN IMAGE. NEURAL NETWORKS OFFER THE PROMISE OF PRACTICAL SOLUTIONS TO COMPUTER-BOUND PROBLEMS. ATLANTIC HAS BEEN PERFORMING LOW LEVEL FEATURE EXTRACTION USING NEURAL NETS, AND HAS BEEN DEVELOPING A FEATURE EXTRACTOR WHOSE GOAL IS THE RELIABLE DETECTION OF PERCEPTUALLY IMPORTANT PHYSICAL FEATURES IN A PARALLEL FASHION. OUR JET-BASED SINGULARITY NETWORK ACHIEVES THIS BY USING FUNDAMENTAL CONSTRAINTS UPON THE RANGE OF POSSIBLE SIGNIFICANT IMAGE FEATURES. THE MANNER THAT THE NETWORK FORMS THESE FEATURES LEADS TO A SIMPLE AND NATURAL SYMBOLIC REPRESENTATION. WE PROPOSE TO EXTEND UPON OUR NEURAL NETWORK WORK IN SINGULARITY EXTRACTION TO INCORPORATE THE INFORMATION FROM SENSORS, SUCH AS PASSIVE INFRARED AND LASER RADAR INTENSITY, WHICH INHERENTLY PRODUCE AREA INFORMATION. IN OUR PROPOSED APPROACH, THIS LOW LEVEL PIXEL ORIENTED DATA WILL BECOME THE FIRST STAGE OF A NEURAL NET PROCESSING HIERARCHY WHICH DEVELOPS PROGRESSIVELY MORE COMPLEX FEATURES FROM THE SENSOR DATA, AND FINALLY USES A NETWORK BASED THREE DIMENSIONAL MODEL TO ACHIEVE THE FINAL IDENTIFICATION OF THE OBJECT. THIS PROCESSING HIERARCHY WILL ALLOW FOR THE FUSION OF INPUT FROM MULTIPLE SENSORS AND ALLOW TARGET IDENTIFICATION WITHOUT DIRECTION CONSTRAINTS WHILE PROVIDING THE CAPABILITY FOR THE IDENTIFICATION OF MULTIPLE OBJECTS IN A SINGLE IMAGE.