Abstract
Artificial Intelligence (AI) has gained prominence in every area over the last two decades at an insurmountable speed. The 4th industrial revolution (Industry 4.0), commonly used to describe the fading of boundaries between the physical and digital worlds, is the primary driving force behind this change, pushed by market forces with the goal of cross-industry automation. Amongst the various applications of AI, computer vision (CV) has gained popularity in road safety, driving a fundamental change in transportation safety using the concept of object detection. How does a machine understand and see the physical world as we see it? To answer this question, we examine CV algorithms that can enable such a process. In this thesis, we examine, primarily, object detection algorithms that utilises Deep Learning (DL) Neural Networks (NN) that have consistently proven to be the optimal method for CV over the last decade. To enable such a process in the area of road safety, we look at factors such as the physical make-up of camera systems that can enable the object detection models to be more aware of the immediate physical surroundings of a vehicle, focusing specifically on fisheye cameras that, though more effective in capturing a physical scene due to its wider field-of-view, will cause distorted images that can decrease the performance of object detection models. We also look at the latency of the overall system, focusing to achieve real-time object detection. Our work will evaluate the performances of current object detection algorithms and the training of a model to achieve such a detection. Finally, we will look to implement such a model on a lightweight, edge device that relies on its on-board computing capabilities (as opposed to most cloud inference models used today).