There exist a number of powerful methods for detecting small, low observable targets with stationary dynamics in image sequences provided by IR and other imaging sensors. However, in real world applications these methods need to be extended to handle maneuvering targets. We demonstrated that the 3D matched filter can be cast into a general framework of optimal spatio-temporal Bayesian filtering as a bank of interacting filters. Target dynamics is modeled by switching multiple models. In contrast to previous studies, we do not assume that the mode jump process is a Markov chain. In particular, we allow the probabilities of jumps to be conditioned on the state variable. We developed a computationally efficient (real time) Track Before Detect algorithm. The algorithm facilitates optimal fusion of sensor measurements and prior information regarding possible threats.
Click here to watch Tracking Before Detection of an agile target in cluttered 2D images. In the movie, the picture on the left-hand side shows the evolution of posterior density of target location on the plane, while on the right-hand side one can see the results of tracking with the use of the maximum posterior estimator. The solid line shows the true trajectory of the target, while the estimated target location is shown by the circles. In this case, the SNR was -3dB, the force of the turn was 7G, and the size of the target was 3x3 pixels. It is seen that the true trajectory was estimated very accurately.