People detection using Siamese networks

Abstract

In modern age computing technologies, detection and tracking of a complex moving target still remains as a major challenge in an unknown environment. Object detection and tracking tasks is becoming more essential and important where it can be used to fight against crimes, terrorism and public safety. One example is the surveillance system. It is often installed in important and crowded places such as shopping malls, hospitals, industrials, military camps etc. Its purpose is to monitor the vicinity for suspicious activities. Recently, deep learning has emerged as one of the powerful platform for tracking and detection tasks as it has been proven that it has the capability to improve the accuracy and performance to effectively detect and track suspects for monitoring systems. By using deep learning in tracking and detection tasks, the risks, costs and effort involved can be reduced. Therefore, training and equipping the person with the essential skills to recognize suspects is no longer needed as deep-learning can also perform the same tasks. When the suspect is tracked and identified using deep learning, the features and information are validated and verified by the computer system. However, due to several factors such as different camera viewpoints, un-clarity of images, occlusions and poor illumination, the accuracy of detecting and tracking of the suspect is reduced. The final year project aims to apply the knowledge of deep learning to solve the challenges faced in human detection and tracking. It is achieved by designing a Siamese architecture which will explore the advantages of reusing features in the deep learning model with the aim to improve its accuracy on detection and tracking tasks. The experimental results will be tabulated and discussed for further improvements in developing a more efficient surveillance system.