Unsupervised learning of background modeling parameters in multicamera systems

Background modeling algorithms are commonly used in camera setups for foreground object detection. Typically, these algorithms need adjustment of their parameters towards achieving optimal performance in different scenarios and/or lighting conditions. This is a tedious process requiring considerable effort by expert users. In this work we propose a novel, fully automatic method for the tuning of foreground detection parameters in calibrated multicamera systems. The proposed method requires neither user intervention nor ground truth data. Given a set of such parameters, we define a fitness function based on the consensus built from the multicamera setup regarding whether points belong to the scene foreground or background. The maximization of this fitness function through Particle Swarm Optimization leads to the adjustment of the foreground detection parameters. Extensive experimental results confirm the effectiveness of the adopted approach.

Research highlights
► Unsupervised optimization of foreground detection parameters.
► Multicamera consensus guides the definition of foreground in an observed scene.
► Particle Swarm Optimization of foreground detection parameters.
► Online, automatically adjustable estimation of foreground detection parameters in varying illumination conditions.
► The proposed approach can be applied to other problems where multiview consensus can be exploited to relax the requirement for ground truth data and/or supervision.