HUMAN TRACKING USING PARTICLE FILTER BASED ON SWITCHING ADAPTIVE/NONADAPTIVE OBSERVATION MODEL

In this paper, we describe a human tracking method based on a particle filter with the switching observation model. During tracking humans in video scenes, occlusions and shape changes of target human often occur. Adaptive observation model in which the model is updated in every frame is effective for shape changes; however, it causes wrong tracking in occlusion scene. To realize robust tracking, switching observation model is introduced. In the method, adaptive observation is basically used, and the update of observation model is stopped when an occlusion occurs. To detect an occlusion, likelihoods are used. The model is applied to sonic datasets, and the effectiveness of the method is verified.