Visual Tracking of Multiple Persons in a Heavy Occluded Space Using Person Model and Joint Probabilistic Data Association

This paper presents a new approach to the problem of image-based tracking of multiple persons in a heavy occluded space using a single camera. The presence of heavy occlusions results in uncertain measurement data. Examples of heavy occlusions are objects which impede the observation of a person, the overlap of multiple persons, etc. This measurement uncertainty can be partially by-passed if the process knows more about a person's expected size, i.e. person model. This way the observed measurement can be corrected using the introduced person model. Also the uncertainty of the measurements will be calculated with this person model. Subsequently, the corrected measurement is used to estimate the person's state (i.e. position and velocity) in the Kalman filter resulting in a more robust tracking. Next, tracking multiple persons jointly implies the need for a data association technique. This paper uses the Joint Probabilistic Data Association (JPDA) filter which calculates the a posteriori probabilities of the measurements having probably originated from the tracked persons. Finally, the approach has been implemented and tested on a single static camera bearing in mind that it will be applied on a mobile camera or robot. The approach presented here will verify whether the use of a single camera, wherefrom only 2D image-based data is gathered, delivers satisfactory tracking results using the Kalman and JPDA filter.