Big Crowd Data Analytics using Bayesian Inference and Deep Learning

It is envisaged that Deep Learning (DL) will be the main approach of the next generation of computer vision research and technology. DL is a branch of Artificial intelligence (AI) based on convolutional neural networks that have shown significant improvements in performance compared to state-of-the-art machine learning (ML) techniques in many domains, ranging from Cardiologists level arrhythmia detection to near human accuracy image classification. Unlike traditional ML approaches which are based on hand engineered features and well defined algorithmic approaches. DL extracts high level features for a given task by processing input data in a manner close to the way the brain process visual and auditory inputs. However, although DL architectures are based on well understood computational units, their choice in a deep architecture is only made via empirical validation, fine tuning of hyper-parameters and the use of big data. More recently, a number of attempts have been made to provide a theoretical justifications and insights into DL successes. Such insights would help building a rigorous theoretical framework for the analysis and synthesis of deep neural architectures and eventually designing more robust architectures with reduced cost and time. This project will focus on proposing breakthroughs of the design of deep learning architectures in the context of computer vision. In particular, the main objective is to develop smart vision sensors for rapid scene categorization and multi-class object detection and tracking in complex dynamic environments. Our goal is to develop deep architectures that transform a simple camera into a smart vision sensor endowed with the capability to detect abnormal behavior, sense dangerous situations, understand hazardous events, identify a familiar face, and locate obstacles. Such technology can be used to alert drivers, assist blind people to navigate freely in crowded areas, guide people affected with dementia towards familiar places, alert law enforcement officers of suspicious activity, assist doctors to make a diagnosis, track persons and objects of interest, or monitor the environment. The project is organized into six linked technical and application work packages that enable a dynamic and a close collaboration between the research team members, with the aim of transforming research ideas into research outcomes in a timely and efficient manner: Technical work packages will develop a new generation technology including DL architectures, multi-class object detection and tracking algorithms. Parallelization of the developed algorithms for real time implementation on heterogeneous hardware will be considered. Application work package will include the integration of the technology developed in the technical work packages into a working software prototype for rapid scene categorization and multi-class object detection and tracking in complex dynamic environments. This project is expected to directly benefit Transport and Communication sectors in the specific context of Doha, within the smart-Doha 2022 perspective. Furthermore, the project is expected to benefit the defense and security industries, enhance law enforcement and military capabilities, improve quality of life of the visually impaired, and thus position Qatar among the leading nations in cutting edge technology. There is significant interest in systems that can detect and track multiple objects across different cameras, recognize crowd behavior, and detect abnormal events, to name a few. The project will advance knowledge, train highly skilled researchers, produce high quality refereed publications, generate IP, and contribute to building a knowledge economy by developing Frontier Technologies.

Hamad Bin Khalifa University (HBKU)