Multiple motion analysis for intelligent video surveillance.

摘要

With the proliferation of camera sensors deployed world widely, video surveillance systems are gradually finding their way into our daily lives. A direct consequence of these technological advancements is the increased demand for intelligent video analysis and understanding techniques.; This dissertation concentrates on the developments of efficient and effective multiple motion analysis techniques that allow automated tracking of multiple targets, which is arguably the most challenging problem and essential component of any intelligent video surveillance systems.; Besides sharing the common challenges faced by visual tracking of single target, including large appearance variations, complex object motions, successful tracking of multiple targets' motions is also confronted by the tremendous difficulties from the theoretical and practical aspects of the problems, such as target occlusions, unknown number of targets, ambiguities of multiple target-tracker associations, high computational demanding, and difficulty of training a target detector.; This dissertation presents several effective and computationally efficient techniques to addressing these challenges: a dynamic Bayesian network formulation for the multiple target tracking with explicit occlusion reasoning; a decentralized framework to multiple target tracking based on Markov network that handles the variable number of targets and copes with the tracker coalescence problem with close to linear complexity; a novel two-layer statistical field model to characterize the large shape variability and partial occlusions for nonrigid target detections, especially pedestrian detections; a component-based appearance tracker based on support vector machines to accommodate the large object appearance variations with the extra appealing capacity of automatically selecting trustworthy components while down-weighting the unreliable occluded components; a novel differential tracking approach based on a spatial-appearance model (SAM) formulation to combine the local appearances variations and global spatial structures enabling the continuous tracking of non-rigid objects that exhibit dramatic appearance deformations, large object scale changes and partial occlusions. Extensive experiments and very encouraging results on both the synthetic and real-world data verified the effectiveness and efficiency of the proposed methods.