Content area
This thesis presents a portfolio of three industrial research and development projects that trace the evolution from conventional machine learning to deep learning approaches in computer vision applications. The work focuses on developing resource-constrained computer vision algorithms for automotive safety and consumer electronics applications between 2010 and 2020.
The first project develops novel feature extraction methods for pedestrian detection in lowlight conditions. The work introduces four methods, namely Weighted Local Binary Pattern (WLBP), Multi-Resolution Local Binary Pattern (MRLBP), Magnitude-Area-Zone-Enhanced Local Binary Pattern (MAZELBP) and Staggered Multi-Scale Local Binary Pattern (SMSLBP). These methods demonstrate improved detection accuracy over existing approaches while maintaining computational efficiency. The work resulted in three US patents, two academic papers and commercial implementation of vehicle safety systems.
The second project addresses driver monitoring applications using a single time-of-flight camera. The project realises human detection by employing SMSLBP. The work also marks a transition to deep learning by employing a shared Convolutional Neural Network (CNN) architecture for skeleton detection and head pose estimation. The system successfully realises a proof of concept and achieves target accuracy while meeting strict automotive requirements for processing speed and resource utilisation.
The third project develops a neural network architecture called Lightweight and Versatile Network (LVNet) for consumer electronics applications. The architecture innovatively increases network nonlinearity through Connection Blocks while significantly reducing computational requirements compared to existing methods. The LVNet enables scene recognition in televisions and event recognition in Blu-ray recorders while securing a US patent. This adoption demonstrates the successful commercialisation of AI capabilities in resource-constrained consumer devices.
The thesis critically analyses the progression across the three projects by examining how they reflect broader computer vision trends and contribute to the field’s advancement. The work demonstrates the successful development and deployment of computer vision algorithms that balance performance with practical constraints. The projects resulted in a technical report, academic papers, patents and commercial implementations. The research and development provide valuable insights into the evolution of industrial computer vision applications and establish frameworks for implementing AI capabilities in resourceconstrained environments.