Object Tracking and Searching in Distributed Camera Networks

Technological advances have created new opportunities in distributed camera networks. One of the main challenges is effective utilization of network-wide knowledge given the constraint of limited network bandwidth and camera node computations. This dissertation addresses the challenge and proposes novel approaches for two basic vision tasks: (1) object tracking and (2) object browsing and searching.;Object tracking is formulated as a global Bayesian estimation problem solved through a distributed Monte-Carlo sampling implementation. At each camera, a particle filter tracker with discriminative appearance model is utilized for image plane tracking. Two novel distributed active fusion schemes are proposed to facilitate camera collaboration. An online jointly-learned discriminative model is used for enforcing appearance consistency by weighting particles from the local tracker. Then object's ground plane motion consistency is enforced by correcting local tracker's particles that deviate from the ground plane estimate from information shared across cameras. The proposed method enables an efficient closed loop interaction between object's local tracking module and the global fusion schemes for robust tracking. Experimental results verify the efficacy of the methodology.;For object browsing/search, we propose a distributed architecture to assist human image analysts to effectively browse and search for objects in a camera network. In contrast to existing approaches that focus on finding global trajectories across cameras, the proposed approach directly models the relationship among raw camera observations. A graph model is proposed to represent tracked objects, their appearance and spatial-temporal relationships. To minimize communication requirements, raw video is processed at camera nodes independently to compute object identities and trajectories at video rate. However, this would result in unreliable object locations and/or trajectories. The proposed graph model captures the uncertainty in these observations by modeling their global relationships, and enables users to query, browse and search the data collected from the network. A graph ranking framework is proposed for the search and retrieval task, and the absorbing random walk algorithm is adapted to retrieve a representative and diverse set of video frames in response to a user query. Experimental results on a wide area camera network are presented.