Integrating statistical and cognitive model for multi-object tracking in realistic scenarios

In this paper, we have addressed a quite researched problem in vision for tracking objects in realistic scenarios containing multifarious situations. We explore cognitive modeling approaches with statistical modeling for tracking objects in contrast to conventional multi-hypothesis and global data association approaches. Our framework comprises of three phases: object detection, integrated cognitive and statistical model, and object tracker. The objects are detected using improved background subtraction with shadow removal technique. Second module is the key to proposed approach and the motivation is to tackle the tracking problem by axiomatizing and reasoning human-tracking abilities with associated weights. An undirected network of detected objects is built in space. Each object contains a unique identity and a data structure of cognitive and statistical attributes whilst satisfying the global constraints of continuity during motion. Consequently, results are linked with Kalman filter based tracker to estimate the trajectories of moving objects. We show that combining cognitive and statistical information gives a straightforward way to interpret and disambiguate the uncertainties due to conicted situations in tracking. The performance of the proposed approach is demonstrated on a set of videos representing various challenges. Besides, quantitative evaluation with annotated ground truth is presented.