Biologically inspired object tracking: A modular approach with distributed particle like sensors

Innovation in computing technology especially in embedded processors capabilities has opened up the boundaries for an enormous potential to facilitate automation in domestic consumer market normally for the sake of added safety and reliability [1][2]. These applications also demand real time processing capabilities for example as required in Surveillance. Although majority of research in computer vision revolves around adopting an analytical approach to this problem, but trackers solely dependent on such algorithms e.g. Kalman filters [3] are not always feasible due to the presence of non-linearity or non gaussian like state space distributions. On the other hand particles based approach [4] requires intensive computing and hence could fail to comply with situations needing real time interventions. This paper is an attempt to highlight such issues and to propose a modular but evolutionary tracking algorithm based on particle swarm optimization [5], distributed sensors in relevant search space and on an algorithm that dynamically changes the complexity based on scene conditions. To reduce the amount of processing a global best search space is first formulated using selective histogram back projection algorithm. The prediction step projects the prior shape contour of the object of interest onto the current frame and by selective initial positioning of the particles forming a swarm explores the search space as in Fig1. Therefore this controlled dynamics inspired by a modular approach to explore relevant search space not only dramatically decreases the required no of particles in our experiments as compared to the traditional particle swarm approach but also proves more robust and thus facilitates real time tracking of an object in video.