Discriminative Appearance Model for Robust Online Multiple Target Tracking
Abstract
Multiple target tracking algorithm faces challenges of occlusion, halt, merge and split of the moving objects. The change in appearance of the moving targets complicates the tracker. Hence the discriminative appearance model is needed for the robust multiple target tracking. This paper incorporates tracking-by-detection approach along with Kalman filter based motion model. The appearance of the target proposed in this paper is modeled based on object's texture features. Phase congruency derived by gray level co-occurrence matrix (GLCM) constitutes the appearance model of the moving object. Thus the proposed tracker is invariant to image illumination and contrast variation. Confidence based data association helps for track management in this paper. The proposed tracker is evaluated on the standard benchmark datasets namely CAVIAR, PETS2009 and ETH. The experimental results of the proposed tracker demonstrate zero error in identity matching when tested on ETH dataset.