Repository logo
 
Loading...
Profile Picture

Search Results

Now showing 1 - 1 of 1
  • QUIS-CAMPI: Biometric Recognition in Surveillance Scenarios
    Publication . Neves, João Carlos Raposo; Proença, Hugo Pedro Martins Carriço
    The concerns about individuals security have justified the increasing number of surveillance cameras deployed both in private and public spaces. However, contrary to popular belief, these devices are in most cases used solely for recording, instead of feeding intelligent analysis processes capable of extracting information about the observed individuals. Thus, even though video surveillance has already proved to be essential for solving multiple crimes, obtaining relevant details about the subjects that took part in a crime depends on the manual inspection of recordings. As such, the current goal of the research community is the development of automated surveillance systems capable of monitoring and identifying subjects in surveillance scenarios. Accordingly, the main goal of this thesis is to improve the performance of biometric recognition algorithms in data acquired from surveillance scenarios. In particular, we aim at designing a visual surveillance system capable of acquiring biometric data at a distance (e.g., face, iris or gait) without requiring human intervention in the process, as well as devising biometric recognition methods robust to the degradation factors resulting from the unconstrained acquisition process. Regarding the first goal, the analysis of the data acquired by typical surveillance systems shows that large acquisition distances significantly decrease the resolution of biometric samples, and thus their discriminability is not sufficient for recognition purposes. In the literature, diverse works point out Pan Tilt Zoom (PTZ) cameras as the most practical way for acquiring high-resolution imagery at a distance, particularly when using a master-slave configuration. In the master-slave configuration, the video acquired by a typical surveillance camera is analyzed for obtaining regions of interest (e.g., car, person) and these regions are subsequently imaged at high-resolution by the PTZ camera. Several methods have already shown that this configuration can be used for acquiring biometric data at a distance. Nevertheless, these methods failed at providing effective solutions to the typical challenges of this strategy, restraining its use in surveillance scenarios. Accordingly, this thesis proposes two methods to support the development of a biometric data acquisition system based on the cooperation of a PTZ camera with a typical surveillance camera. The first proposal is a camera calibration method capable of accurately mapping the coordinates of the master camera to the pan/tilt angles of the PTZ camera. The second proposal is a camera scheduling method for determining - in real-time - the sequence of acquisitions that maximizes the number of different targets obtained, while minimizing the cumulative transition time. In order to achieve the first goal of this thesis, both methods were combined with state-of-the-art approaches of the human monitoring field to develop a fully automated surveillance capable of acquiring biometric data at a distance and without human cooperation, designated as QUIS-CAMPI system. The QUIS-CAMPI system is the basis for pursuing the second goal of this thesis. The analysis of the performance of the state-of-the-art biometric recognition approaches shows that these approaches attain almost ideal recognition rates in unconstrained data. However, this performance is incongruous with the recognition rates observed in surveillance scenarios. Taking into account the drawbacks of current biometric datasets, this thesis introduces a novel dataset comprising biometric samples (face images and gait videos) acquired by the QUIS-CAMPI system at a distance ranging from 5 to 40 meters and without human intervention in the acquisition process. This set allows to objectively assess the performance of state-of-the-art biometric recognition methods in data that truly encompass the covariates of surveillance scenarios. As such, this set was exploited for promoting the first international challenge on biometric recognition in the wild. This thesis describes the evaluation protocols adopted, along with the results obtained by the nine methods specially designed for this competition. In addition, the data acquired by the QUIS-CAMPI system were crucial for accomplishing the second goal of this thesis, i.e., the development of methods robust to the covariates of surveillance scenarios. The first proposal regards a method for detecting corrupted features in biometric signatures inferred by a redundancy analysis algorithm. The second proposal is a caricature-based face recognition approach capable of enhancing the recognition performance by automatically generating a caricature from a 2D photo. The experimental evaluation of these methods shows that both approaches contribute to improve the recognition performance in unconstrained data.