OCCLUSION MODEL FROM HUMAN INTERACTION ANALYSIS FOR TRACKING MULTIPLE PEOPLE

摘要

In this work we investigate the problem of tracking mul-rntiple interacting people under uncontrolled stationary en-rnvironments for intelligent surveillance applications. Thisrndomain is very challenging since the clothing appearancernchanges of the people over time make difficult the temporalrnassociation of their identities. The problem is emphasizedrnwhen individuals move close to each other, are occluded,rnor abruptly change their trajectories.rnWe propose a tracking graph that models spatial andrntemporal relationships among people in order to predict andrnresolve partial and total occlusions. When a total occlusionrnevent occurs, the model generates three possible hypothe-rnses about the location of the occluded person according torna human interaction analysis. This model is able to detectrnfalse positives and false negatives in the detection measure-rnments and it can also estimate the location of missing orrnoccluded people.rnOur approach was evaluated on benchmark sequencesrnand results show how it outperforms state-of-the-art algo-rnrithms even in the presence of long periods of occlusion.