| This dissertation addresses the problem of tracking human targets in complex, cluttered environments. This novel approach, called the Cluster Tracker, concentrates on insulating and tracking small features rather than the whole image as more conventional trackers do. We found that this approach exhibits higher tracking reliability in heavily cluttered environments due to its ability to maintain track lock on objects even when they are partially obscured. The core of the method is based on an Elastic Matrix framework that supports a flexible structured model of the target. This approach allows the tracker to follow the deformations of the target's body and to estimate feature locations when occluded.; Tracking humans poses a special challenge since humans continuously change shape as we walk, turn and interact with our environment. Building a model of an unconstrained target without a priory knowledge is exceedingly difficult, so we chose to build a non-model structured approach by automatically selecting regions of the target image that exhibit high trackability and coherently tracking them as a group. The proposed approach accurately tracks polymorphic targets exhibiting rigid and non-rigid motion without having any prior knowledge of the target's structure or its environment. The method is also computationally efficient and able to cope with dynamic backgrounds and severe foreground occluding environments. |
