A hybrid blob- and appearance-based framework for multi-object tracking through complex occlusions

Static and dynamic occlusions due to stationary scene structures and/or interactions between moving objects are a major concern in tracking multiple objects in dynamic and cluttered visual scenes. We propose a hybrid blob- and appearance-based analysis framework as a solution to the problem, exploiting the strength of both. The core of this framework is an effective probabilistic appearance based technique for complex occlusions handling. We introduce in the conventional likelihood function a novel 'spatial-depth affinity metric' (SDAM), which utilises information of both spatial locations of pixels and dynamic depth ordering of the component objects forming a group, to improve object segmentation during occlusions. Depth ordering estimation is achieved through a combination of top-down and bottom-up approach. Experiments on some real-world difficult scenarios of low resolution and highly compressed videos demonstrate the very promising results achieved.