Adaptive Context-Based Systems for Face Recognition in Video Surveillance

点击翻译按钮获取中文摘要

Still-to-video face recognition (FR) is an important function in video surveillance, where faces captured across a network of video cameras are matched against reference stills images. Currently, the performance of state-of-the-art systems is severely affected by the multiple non-stationary operational environments, defined by variations in, e.g., pose, illumination and camera viewpoint. Moreover, since an individual is typically enrolled using one or few reference stills captured during enrolment, face models are often poor representatives of the faces to be recognized during operations. ***The main objective of this research program is to investigate and develop new still-to-video FR systems for accurate spatio-temporal recognition of individuals of interest across multiple cameras. To compensate for the limited number of labeled reference stills available during enrolment, the system's facial models will adapt based on contextual information from its operational domain (Od) in order to sustain a high level of performance. In particular, this research relies on a bank of unlabeled (non-target) video trajectories from each different camera in a surveillance network. For each Od, this these systems will allow to: (1) model the distribution of capture conditions in facial trajectories according to feature and ROI quality spaces, (2) generate accurate detectors for each enrolled individual using instance- and feature-based methods for domain adaptation, (3) generate synthetic reference ROIs, (4) weight and integrate the response of each detector according to capture condition, and (5) perform efficient self-update of detectors driven by changes in capture conditions. This program is organized according to 3 axes: ***1. Modeling the capture conditions densities from Od trajectories.***2. Dynamic ensembles of exemplar SVMs that exploit domain adaptation for accurate detection of individuals. New techniques will be developed to assess classifier competence, self-update, and generate synthetic stills.***3. Extended sparse representation classifiers that use domain adaptation for accurate detection of individuals. New techniques will be developed to compress, synthetically generate, and specialize dictionaries to an Od.***Innovative techniques developed in this program will contribute to the emergence of context-sensitive FR technologies for watchlist screening, and for several other visual monitoring and recognition applications. As such, this research will remain relevant in the areas of pattern recognition, evolutionary computing and computer vision. Although government, military and law enforcement remain the principal users of video surveillance technologies, the recent emergence of many commercial applications and consumer devices will also offer excellent opportunities for partnering with companies, training HQP, and encourage future grant applications.**