Tracking and classifying 3D objects from TV pictures-a neural network approach

The authors introduce two specialized neural networks, one for tracking and another for classification of a 3-D moving object from TV pictures. Their structures follow directly from the analysis of the corresponding functions. It is shown that this approach gives equally good results in tracking tasks as the optimal Kalman filtering procedure, and that it even overcomes classification results obtained with Fourier transformations of an object´s contours. With parallel processing real time applications seem feasible. The input images are taken sequentially from digitized TV pictures showing the 3-D object´s movement along its trajectories. It is the responsibility of the tracking function to make the changes of the window properly. It is shown that with the proposed tracking neural network the same quality of tracking can be achieved as with the Kalman filtering method. The classification of moving objects performed by the classification neural network also gives encouraging results.