STUDY OF MULTI-CUE OBJECT TRACKING IN VIDEO SEQUENCES

Abstract

Multi-cue object tracking is a challenging eld of computer vision. In particular, the challenges originate from environmental variations such as occlusion, similar background and illumination variations or due to variations in target's appearance such as pose variations, deformation, fast motion, scale and rotational changes. In order to address these variations, a lot of appearance model have been proposed but developing a robust appearance model by fusing multi-cue information is tedious and demands further investigation and research. It is essential to develop a multicue object tracking solution with adaptive fusion of cues which can handle various tracking challenges. The goal of this thesis is to propose robust multi-cue object tracking frameworks by exploiting the complementary features and their adaptive fusion in order to enhance tracker's performance and accuracy during tracking failures. A real-time tracker using particle lter under stochastic framework has been developed for target estimation. The inherent problems of particle lter namely, sample degeneracy and impoverishment have been addressed by proposing a resampling method based upon meta-heuristic optimization. In addition, an outlier detection mechanism is designed to reduce the computational complexity of the developed tracker. A robust tracking architecture has been proposed under deterministic framework. Fragment-based tracker with a discriminative classi er has been designed that can enhance tracker's performance during dynamic variations. Periodic and temporal viii update strategy is employed to make tracker adaptive to changing environment. Extensive experimental analysis has been performed to prove the e ectiveness of the developed tracking solution. Multi-stage tracker based on adaptive fusion of multi-cue has been developed for multi-cue object tracking. The rst stage of target rough localization improves the accuracy of tracker during precise localization. In the appearance model complementary cues are considered to handle illumination variations and occlusion. Classi er mechanism and fragment based appearance model are proposed to improve the tracker's accuracy during background clutters and fast motion. Experimental validation on multiple datasets validates the performance and accuracy of the proposed tracker.