New approaches for dim target detection and clutter rejection

This dissertation presents a new method for clutter rejection and dim target track detection from infrared (IR) satellite data using neural networks. A method referred to as "high order correlation method" is developed which recursively computes the spatio-temporal cross-correlations between data of consecutive scans. The implementation of this scheme using a connectionist network is also proposed. Several important properties of the high order correlation method are established which indicate that the resultant filtered images capture all the target information. Simulation results using this approach show at least 93% clutter rejection under moderate clutter density. Further improvement in the clutter rejection is achieved by modifying the high order correlation method to incorporate the target motion dynamics. The implementation of this "modified high order correlation" using a high order neural network architecture is also developed. Simulation results indicate at least 97% clutter rejection rate for this method.; To test the performance, experimental studies of this modified high order correlations are conducted under various scenarios which include: multiple target detection, continuous mode operation, various background clutter densities, and detection using variable number of scans and order of correlation. This algorithm performs very well even under many difficult operating environments.; A new scoring process is developed to improve the discrimination ability of the modified high order correlation scheme by employing velocity and curvature information. This scoring process is then used to identify each individual track in the scene by using the properties of the modified high order correlation method. This modification not only significantly improves the clutter rejection under very dense clutter environment, but also increases the feasibility of using the modified high order correlation method for other areas such as data association, target classification and tracking.; The features and effectiveness of several conventional approaches are discussed and some details are given on probabilistic data association, three-dimensional (3-D) filtering and neural networks-based approaches. The methods developed in this dissertation are also benchmarked against the frequency domain 3-D filtering scheme. This comparison revealed that the proposed schemes completely outperform this method.