Multiple object tracking on static surveillance video using field-based prediction information in MPEG-2 video

We present new algorithms to detect and track multiple moving objects in a static surveillance video that uses MPEG-2 compression. The algorithm detects moving objects based on the location of field-based motion estimated macroblocks in inter-prediction frames, and tracks their movements along video scene by using an extension of Kalman filter that uses the objects´ velocity information. Experiments show that the algorithm outputs highly accurate results in various scenarios.