NEURAL NETWORK-BASED CLASSIFICATION OF VEHICLES FROM LASER RADAR DATA

HIGHLY RELIABLE VEHICLE CLASSIFICATION VIA COHERENT RADAR VIBRATION DATA HAS BEEN DEMONSTRATED USING PARAMETRIC CLASSIFIER TECHNIQUES. HOWEVER, PARAMETRIC CLASSIFIER DEVELOPMENT REQUIRES EXTENSIVE, EXPERT HUMAN DIRECTION. IT BECOMES IMPRACTICAL WHEN THE CLASSES ARE COMPLEX, AS WHEN THERE ARE MANY OPERATIONAL AND ENVIRONMENTAL CONDITIONS AND WHEN THE NUMBER OF CLASSES IS LARGE. NEURAL NET FORMULATIONS, PARTIULARLY FEEDFORWARD NETWORKS, HAVE SHOWN MUCH PROMISE IN CHARACTERIZING COMPLEX CLASSES. HOWEVER, THE OPTIMUM NUMBER OF NODES CAN BE DETERMINED ONLY BY TRIAL AND ERROR. TRAINING SUCH NETWORKS BY ITERATIVE METHODS CAN BE PROHIBITIVELY EXPENSIVE. INTERPRETING THE FEATURES DISCOVERED BY A NETWORK CAN BE HIGHLY PROBLEMATICAL. WE DESCRIBE A DIRECT METHOD NET CONSTRUCTION, BASED ON POTENTIAL FUNCTION CLASSIFIER THEORY, WHICH IS MORE GENERAL THAN ANY NOW IN USE. GREAT SUPERIORITY IN EFFICIENCY AND ACCURACY OVER NEAREST NEIGHBOR CLASSIFICATION IS DEMONSTRATED IN TESTS ON SYNTHETIC DATA. WE PROPOSE TO CONSTRUCT SUCH A NETWORK AND EVALUATE ITS PERFORMANCE ON A DATA SET WHICH HAS BEEN USED TO CONSTRUCT AND TEST A PARAMETRIC CLASSIFIER. AN ENCODING OF THE SPECTRAL DATA AS FEATURES APPROPRIATE FOR INPUT TO THE NET IS DISCUSSED. THESE FEATURES SHOULD EXTRACT ALL OF THE SIGNIFICANT SPECTRAL INFORMATION AND MAINTAIN A STABLE CONFIGURATION IN THE PRESENCE OF NOISE.