On-line Learning of Motion Patterns using an Expert Learning Framework

Tracking uncertain mobile objects such as humans and vehicles is an important problem in computer vision, robotics, and geo-spatial visualization. As the name suggests, predictor-corrector tracking is performed in two steps -prediction and correction. Prediction steps have typically utilized a-priori motion model most common of which is a uniform motion model. In this work, we apply an expert learning framework for on-line prediction and learning the motion of an uncertain mobile object. We define a number of probabilistic experts, each of which predicts the future position of the object with some uncertainty and then combine the predictions of all the experts to produce an estimate of the object's location. Individual experts predictions are weighted adaptively depending on their performance. We show that this adaptive combination is powerful when there are changes in the pattern of the object's motion. Results of our algorithm are compared with linear extrapolation and the best off-line expert predictions. We have tested our algorithm with synthetic data using uniform and non-uniform patterns as well as real data acquired using GPS equipment in presence of intermittent and highly erroneous data.