Human re-identification using deep learning algorithm
Low, De Wei
Date of Issue2017-05-19
School of Electrical and Electronic Engineering
The rise of external threats from terrorism attacks upon infrastructures and businesses has led to the widespread usage and reliance on strong and robust surveillance and security systems. Human Re-identification is defined as the process of reidentifying the person of interest through different camera or surveillance systems. Human Re-identification has been widely used in surveillance systems worldwide to allow better surveillance and tracking of suspicious identities through different surveillance and camera systems. Automated human re-identification systems reduce the need for manpower and also the time and cost required to train them. However automated human re-identification remains a challenge as the application of these systems is at densely populated area like the airport, shopping malls and places of interest. This gives rise to problems such as occlusion of the person of interest which prevent the system from accurately identify the person of interest. Poor resolution of the different cameras used and low lighting and illumination of surroundings will also affect the robustness and accuracy of the human re-identification system. This final year project seeks to apply and extend the use of deep learning neural network to solve the challenges faced in human re-identification. This project will also be exploring the usage of different colour spaces and the number of colour channels as features to be learned by the neural network. The project will be spilt into three main phases. The first phase will be conducted by using different colour spaces like hue and grayscale from image pairs to form 2- channel colour space positive and negative images. The prepared images will be fed into benchmark convolution neural network architecture for model training. The second phase of the project will be exploring the usage of more colour channels as input to the convolution neural network. The second phase will use the RGB components of the paired images to form a six channel RGB image. The third phase will use HSV colour space components as paired images to form a six-channel HSV to be fed into the convolution neural network for learning. Results collected from the three main phases will be discussed.
Final Year Project (FYP)
Nanyang Technological University