A learning network for adaptive target tracking

An adaptive scheme for detecting and tracking moving objects under noisy, dynamic conditions is presented. This scheme integrates the adaptiveness and incremental learning characteristics of neural learning networks with intelligent reasoning and process control. A multineuron learning network is used as a tracked target/clutter filter to work with a moving object tracker. Each neuron forms a surface in the multidimensional feature space, and multiple surfaces are used to enclose the target distribution. Each surface takes care of a subset of clutter. A dynamic network construction scheme that dynamically determines the network architecture by incrementally creating new neurons and deleting redundant neurons in the network is developed. The algorithm is benchmarked against the Widrow-Hoff algorithm for four different sets of forward looking infrared data. In all cases, the algorithm performed better in terms of reduced misclassifications. >