Motion priors for multiple target visual tracking

This article describes an original strategy for enhancing current state-of-the-art trackers through the use of motion priors, built as data-driven probabilistic motion models for moving targets. Our priors have a simple form and can replace advantageously more traditional models, such as the constant velocity or constant acceleration models, that are of common use in visual tracking systems, but that are also prone to fail in handling critical scene-related constraints on the targets motion. These priors are learned based on local motion observed in the video stream(s) and, given that the obtained representation may be incomplete and noisy, we regularize it in a second phase. The hybrid discrete–continuous motion priors are then used within two classical target tracking approaches: (1) as a sampling distribution in a particle filter framework and (2) as a weighting prior in a detection-based framework. For both tracking schemes, we present promising results with our motion prior approach, on classical benchmark datasets from the visual surveillance tracking literature.