ATR for dual-band forward-looking infrared images

We discuss the optimisation of the fusion of data captured from two wavebands, 3-5 μm and 3-12 μm, from digital forward-looking infrared (FLIR) image sequences in order to recognise vehicles. A single-waveband system has been developed previously (Toulson and Boyce 1996), that utilises two successive neural nets to first recognise and then, after a tracking-stage, to confirm the recognition of vehicles within a sequence of images. Tracking is normally performed entirely in terms of the spatial trajectories of the detections through time and is reliant on a set of consistent detections for the process to operate efficiently. This approach neglects the use of the signals in successive frames as a method of providing enhanced data association. We have implemented a new method for tracking that integrates a Kalman-filter/data-association method with a correlation/centroiding approach