Multi Person Tracking with a Multi Hypothesis Approach for Ambiguous Assignments
Multi-person tracking is often solved with the tracking-by-detection paradigm, in that a distance measure is calculated for each possible track-detection assignment. Then, the sum of distances of all assignments has to be minimized, for which mostly the Hungarian method is used. Wheareas it is easy to design a distance measure that can clearly indicate the correct assignments in sequences with sparse person distributions, the distances of some assignments can be very similar in crowded scenes, where multiple persons share similar spatial positions and appearances. As a consequence, wrong assignments are inescapable, harming the tracking performance. In contrast of executing all assignments simultaneously, no matter if they are clear or ambiguous, this work treats ambiguous assignments with similar distances separately following a multihypothesis approach, updating the hypotheses until the assignment task is clear again. To determine which assignments are considered ambiguous, a method that compares the entries in the distance matrix of track-detection assignments is introduced. No further information next to the distance matrix is needed, which makes the proposed approach applicable to any tracking-by-detection based method. Experimental results show that the separate treatment of ambiguous assignments can improve the tracking performance in crowds and thus is a promising research directory.