Neural network-based detection and tracking of maneuvering targets in clutter for radar applications.

Abstract

Until the recent past, almost all proposed methods for detection and tracking of maneuvering targets in clutter have followed the algorithmic path. For most multi-target tracking problems, however, the algorithmic approach generally requires a speed and a degree of parallelism which is far beyond the capabilities of available computational resources. This dissertation investigates the development of neural network-based methods for detection and tracking of maneuvering targets in clutter background and focuses on three major operations required for this overall task. A detection scheme is developed by utilizing the pattern classification ability of a trained neural network which helps in a better representation of the clutter and the targets. Utilizing the mapping property of neural networks, a higher probability of detection is achieved while preserving a constant rate of false alarm. The second unit is a Moving Target Indicator (MTI) which is trained through examples in order to integrate a series of noisy radar pulses and provide estimates of target radial velocity. For the problem of tracking a maneuvering target, conventional algorithms employ a Kalman filter which provides estimates of the target position and velocity. While a Kalman filter is the most powerful linear estimator for continuous random variables, it may fail to converge in the presence of sharp measurement discontinuities which may be caused by clutter or sudden target maneuvers. A multilayer feedforward neural network in conjunction with a Kalman filter can better resolve the discontinuity in the measurement sequence. In the new approach proposed here, a neural network is trained to provide an on-line estimate of the necessary artificial noise components which will help neutralizing the corresponding bias in Kalman filter estimates of target kinematic parameters.