Object Tracking via Multi-region Covariance and Particle Swarm Optimization

In this paper a particle swarm optimization based algorithm for object tracking in surveillance videos is proposed. Given the estimate of the object state, the particles are drawn from a Gaussian distribution in order to cover the promising object locations. The particle swarm optimization takes place afterwards in order to concentrate the particles near the true state of the object. The optimization aims at shifting the particles towards more promising regions in the search area. The region covariance is utilized in evaluation of the particle score. The object template is represented by multiple object patches. Every patch votes for the considered position of the object undergoing tracking. Owing to robust combining of such patch votes the object tracker is able to cope with considerable partial occlusions. A tracking algorithm built on the covariance score can recover after substantial temporal occlusions or large movements. Through the usage of multi-patch object representation the algorithm posses better recovery capabilities and it recovers earlier. Experimental results that were obtained in a typical office environment as well as surveillance videos show the feasibility of our approach, especially when the object undergoing tracking has a rapid motion or the occlusions are considerable. The resulting algorithm runs in real-time on a standard computer.