Abstract/Summary

This thesis investigates the usage of Siamese networks across three surveillance and security tasks for land border security. Siamese networks (also known as a twin-pair network) are a layout of neural networks that contain a segment that contains duplicated architecture and configuration parameters for feature extraction of two inputs, combining the outputs into one vector for comparison in a final set of layers to produce a similarity score. The effectiveness of multiple architectures of Siamese networks crafted from multiple generations of Convolutional Neural Networks and Residual Neural Networks are examined for side-profile vehicle classification and Differential Morphing Attack Detection (D-MAD), and with a novel architecture for trajectory similarity analysis. The challenging domain of automated vehicle classification from pole-mounted roadway cameras from side-profile views is evaluated. Three Siamese networks based on existing non-Siamese architectures are proposed and compared against five existing methods on a novel and published dataset. The evaluation undertaken shows that the residual based Siamese network is able to outperform other state of the art methods on datasets with a small number of classes. An end-to-end Siamese trajectory network framework is proposed for the purpose of trajectory similarity analysis in surveillance tasks. A deep feature auto-encoding network is used as part of a discriminative Siamese architecture to perform trajectory similarity analysis. The effectiveness of this method is evaluated on four challenging public real-world datasets containing both vehicle and pedestrian targets, and compared with five existing methods. The proposed method outperforms the existing methods on three of the four datasets. Face morphing attacks pose an increasingly severe threat to automatic face recognition systems in border control environments. Three Siamese architectures built up from multiple generations of non-Siamese Convolutional and Residual Neural Networks for D-MAD are proposed, showing the effectiveness of these networks against a pre-established Convolutional architecture for Single-image Morphing Attack Detection (S-MAD). The residual network based architecture outperforms representative convolutional architectures from the literature, with the Siamese D-MAD architecture able to outperform its S-MAD variant.