Abstract:
This thesis undertakes a comprehensive exploration of object detection technologies with a particular emphasis on their application in autonomous driving and related application settings. The research delves into 2D and 3D object detection methods, offering insights into their evolution, current state, and potential future developments. The thesis provides an extensive literature review in the initial chapters, charting the progress from 2D to 3D object detection techniques. It scrutinises the advancements in 2D object detection, focusing on the YOLOv8 architecture, a state-of-the-art model known for its real-time detection capabilities. The discussion encompasses various evaluation metrics, such as IoU and its variants, mAP and the F1 score. The subsequent discussion follows on loss functions, alongside an in-depth tutorial on the architecture and components of YOLOv8. The application of this model in Indian driving scenarios is also explored, offering a unique perspective on its performance in diverse environmental settings. A significant focus of the research is identifying and mitigating errors in current state-of-theart 3D detection models. This aspect is critical for enhancing the safety and reliability of autonomous vehicles, particularly under varying and challenging conditions. The thesis sets forth a problem statement to systematically categorise specific errors inherent in these models and develop strategies to address them effectively. Moving forward, our thesis will delve deeper into the research surrounding 3D object detection algorithms, error analysis, and active learning methods. Our ultimate goal is to enhance object detection models to meet the rigorous demands of autonomous driving. The main focus of our research will be to identify effective techniques that can improve the accuracy and efficiency of these models.