Application of Convolutional Neural Networks in Object Detection, Re-identification, and Recognition
Abstract
This thesis investigates the effective deployment of deep Convolutional Neural Networks (CNNs) architectures in two different application areas for security and surveillance purposes, namely person re-identification and aerial, large-scale, object detection and recognition, in which the data capture sources are cameras within CCTV systems and on-board drones, respectively.
First of all, person re-identification is of significance importance in video surveillance applications and remains an open research problem to-date. A significant research effort has been focused worldwide in this area, where initially traditional machine learning approaches were used as the technology that underpins the person re-identification, but in the recent past with the advent of significant developments in Deep Neural Network (DNN) technology, several efforts have been made in using DNNs in person re-identification. Unfortunately, all such attempts have been limited to using a selected CNN in person re-identification associated with a single dataset, mostly those available in the public domain. Furthermore, such attempts are very limited in their analysis of the potential to optimise such networks for use within a given dataset. In this thesis we therefore conduct a comprehensive investigation on the effectiveness of the use of the state-of-the-art CNN meta-architectures, AlexNet, VGGNet-16, ResNet and Inception V4, in seven benchmark person re-identification datasets. The impact of tuning the learning rate of the networks and the optimization methods, are evaluated, with the aim of optimising the object re-identification accuracy. This research proves the significance of the network hyper-parameter optimisation besides the importance of data preparation, training, testing and evaluation to obtain the optimum performance of a given network when applied to a specific dataset. The rigorous investigations carried out in this thesis and the conclusions thus made closes a significant current research gap in both the effective use of DNN technology, generally in any application and specifically in person re-identification.
The second part of this thesis focuses on the effective use of DNN technology in efficient object detection, in drone-based imagery. In this modern era, drones facilitate accessing images in challenging environments and scanning large ground/terrain areas in a minimum time, which enables many new applications to be established based on automatic computer-based analysis of drone imagery. As drones are typically flown at high altitudes in order to facilitate coverage of large areas within a short time, the captured object size tends to be very small and mostly very low in resolution, thus poses a significant challenge in using drone-imagery for object detection and recognition. The use of traditional machine learning approaches for object detection in drone imagery will thus be very restrictive. However, the latest developments in DNN technology provides hope for the development of object detectors for drone imagery. Unfortunately, the focus at present of this new technology is analysing video footage captured by CCTV or hand-held cameras, where the view angle is significantly different to the view angle of drones. Hence DNN models developed for CCTV or handheld cameras/systems cannot be used in detecting objects in drone-imagery. Such networks need effective re-training on drone imagery. Therefore, this thesis focuses on the effective use of DNN technology in object detection in drone imagery. Effective data capture and preparation, the importance of optimising networks via hyper-parameter tuning, effective network learning strategies, testing and evaluation approaches are rigorously analysis in this thesis. The above is presented in relation to specific object detection tasks in drone imagery, namely, the detection of p