UMass Machine Learning and Friends Lunch | Main / Object Interaction Analysis By Estimating Statistical Dependence

We show how statistical dependence estimation underlies the problem of detecting and describing interactions between moving objects, such as one pedestrian following another. Object interaction analysis is crucial for automated visual surveillance of the behavior of multiple objects. The main idea is that the behavioral structure of object interaction implies a corresponding structure of statistical dependencies in the object motion. We extend previous work, which assume that the interaction structure is known, to inferring the interaction by estimating dependence. Experimental results demonstrating the approach on simulated and real data are shown. Statistical dependence estimation is also useful in other applications such as matching across multiple sensors and audio/video association.