Robust Video Object Tracking in Distributed Camera Networks

摘要

We propose a robust video object tracking system in distributed camera networks. The main problem associated with wide-area surveillance is people to be tracked may exhibit dramatic changes on account of varied illuminations, viewing angles, poses and camera responses, under different cameras. We intend to construct a robust human tracking system across multiple cameras based on fully unsupervised online learning so that the camera link models among them can be learned online, and the tracked targets in every single camera can be accurately re-identified with both appearance cue and context information. We present three main parts of our research: an ensemble of invariant appearance descriptors, inter-camera tracking based on fully unsupervised online learning, and multiple-camera human tracking across non-overlapping cameras.;As for effective appearance descriptors, we present an appearance-based re-id framework, which uses an ensemble of invariant features to achieve robustness against partial occlusion, camera color response variation, and pose and viewpoint changes, etc. The proposed method not only solves the problems resulted from the changing human pose and viewpoint, with some tolerance of illumination changes but also can skip the laborious calibration effort and restriction.;We take an advantage of effective invariant features proposed above in the tracking. We present an inter-camera tracking method based on online learning, which systematically builds camera link model without any human intervention. The aim of inter-camera tracking is to assign unique IDs when people move across different cameras. Facilitated by the proposed two-phase feature extractor, which consists of two-way Gaussian mixture model fitting and couple features in phase I, followed by the holistic color, regional color/texture features in phase II, the proposed method can effectively and robustly identify the same person across cameras.;To build the complete tracking system, we propose a robust multiple-camera tracking system based on a two-step framework, the single-camera tracking algorithm is firstly performed in each camera to create trajectories of multi-targets, and then the inter-camera tracking algorithm is carried out to associate the tracks belonging to the same identity. Since inter-camera tracking algorithms derive the appearance and motion features by using single-camera tracking results, i.e., detected/tracked object and segmentation mask, inter-camera tracking performance highly depends on single-camera tracking performance. For single-camera tracking, we present multi-object tracking within a single camera that can adaptively refine the segmentation results based on multi-kernel feedback from preliminary tracking to handle the problems of object merging and shadowing. Besides, detection in local object region is incorporated to address initial occlusion when people appear in groups.