On Time-varying Size Estimation of Visual Target by State Estimation via Particle Filter

Likelihood design in state space modeling for visual target tracking in a manner of track-before-detect via particle filter may cause problem for estimating size of the target to be tracked. Specifically, color histogram matching is a typical case of the likelihood where two histograms, memorized histogram beforehand applying the tracking procedure and captured histogram for each particle, are compared based on Bhattacharyya distance. In such a case, the likelihood has scale invariant characteristic, so it causes size estimation to fail. Although color histogram matching is old-fashioned method, this kind of phenomena on scale invariant may occur in the recent methods based on deep neural networks, and so on. To circumvent the above difficulty, this paper proposes an idea to incorporate the ratio of target-likelihood and background-likelihood powered by the number of pixels of the target region. Note that the idea has been inspired by separable likelihood appeared in multiple target tracking scenarios. Simulation experiments over the synthetic video images demonstrate performance of the proposed idea with comparison among some variants on likelihood designs.