Motion analysis with recurrent neural nets

Abstract

218zVisual tasks such as the interpretation of cell images (Psarrou and Buxton, 1993) and the recognition of moving vehicles require to track objects along their trajectory and to predict their future position in their environment. It was noted that objects move purposely in an environment and effective prediction on their trajectories can be achieved by modelling the spatio-temporal regularities associated with their moving purposes with visually augmented hidden Markov Models (Gong and Buxton, 1992). Temporal prediction and recognition require (a) a short-term memory that retains aspects of the input sequence relevant to prediction and recognition, (b) the specification of a function that combines the current memory and the current input in order to form a new temporal context (Mozer, 1993), (c) to identify and learn the regularities from the temporal sequence.