We propose a method for multi-object tracking, where interaction between object hypotheses plays a key role. In our method, we generate hypotheses for an object in the image and represent them as certainty distributions in the model-parameter space. We then propagate and reform hypotheses over time in turn. In addition, we bring about interaction between hypotheses to eliminate the hypotheses denoting false positives and, at the same time, to maintain the hypotheses denoting objects. This allows us to identify the one-to-one correspondence between hypotheses and objects in the image. Consequently, the system tracks multiple objects stably even if occlusions occur and the number of objects in the image changes during tracking. Experimental results show the effectiveness of our method.
展开▼