Context-Aware Activity Forecasting

In this paper, we investigate the problem of forecasting future activities in continuous videos. Ability to successfully forecast activities that are yet to be observed is a very important video understanding problem, and is starting to receive attention in the computer vision literature. We propose an activity forecasting strategy that models the simultaneous and/or sequential nature of human activities on a graph and combines that with the interrelationship between static scene cues and dynamic target trajectories, termed together as the ‘activity and scene context’. The forecasting problem is then posed as an inference problem on a MRF model defined on the graph. We perform experiments on the publicly available challenging VIRAT ground dataset and obtain high forecasting accuracy for most of the activities, as evidenced by the results.