Using neural network to learn spatial-temporal models for moving deformable objects tracking

We investigate the effectiveness of using neural networks for multivariate time series prediction on data from a practical machine vision domain. The data, taking the form of eight short sequences of compact state vectors, are descriptions of typical walking behaviors of a pedestrian in an outdoor scene. The prediction of the changes in the state vector based on a bank of neural network models opens a new way to track the motion of the pedestrian within the scene, showing some unique features. The particulars of this data set are analyzed, the emphasis has been laid on the design of network architecture and the formulation of training patterns to accommodate the spatial-temporal variations of the problem. The prediction results of the models trained using the noise regularization method have been obtained with great success.