Fusion of region and point-feature detections for measurement reconstruction in multi-target Kalman tracking

Object tracking in 2D video surveillance image data is one of the key needs for many follow-up operations such as object classification or activity recognition. In scenes with multiple objects crossing each other's way, there is a high potential for split and merge detections disturbing the tracking process. In these situations, it is helpful or even necessary to reconstruct the object-related measurements to support tracking approaches such as Kalman or Particle Filter. We present a way of fusing three different detection approaches taking benefit from their specific advantages to reconstruct measurements, if a split or merge situation is recognized. The resulting split and merge handling shows better results than using each detection approach individually without fusion. Furthermore, the tracking process is fast with a computation time less than one millisecond per image. Experimental results are given in example video scenes of an infrared camera located on a buoy for maritime surveillance.