Temporal validation of Particle Filters for video tracking

We present an approach for determining the temporal consistency of Particle Filters in video tracking based on model validation of their uncertainty over sliding windows. The filter uncertainty is related to the consistency of the dispersion of the filter hypotheses in the state space. We learn an uncertainty model via a mixture of Gamma distributions whose optimum number is selected by modified information-based criteria. The time-accumulated model is estimated as the sequential convolution of the uncertainty model. Model validation is performed by verifying whether the output of the filter belongs to the convolution model through its approximated cumulative density function. Experimental results and comparisons show that the proposed approach improves both precision and recall of competitive approaches such as Gaussian-based online model extraction, bank of Kalman filters and empirical thresholding. We combine the proposed approach with a state-of-the-art online performance estimator for video tracking and show that it improves accuracy compared to the same estimator with manually tuned thresholds while reducing the overall computational cost.