Improved Perceptions of Autonomous Vehicles In Urban Mobility Environment Through Deep learning Object Detection Algorithms

Abstract

Autonomous Vehicles (AVs) or self-driving cars have become the most exciting area of research in the field of transportation in the last decade. Implementation of AVs in real-world requires further investigation especially in the contexts of safety and comfort of vulnerable road users (pedestrians and cyclists). This thesis focusses on exploring the impact of autonomous vehicles (AV) and their abilities from a computer-vision-based perspective. The thesis initiates with an extensive literature review, delving into the current landscape of detecting and embracing AV. The primary objective is to explore the utilization of deep learning methodologies for detection of vulnerable road users. Following the review of the state-of-the-art, a technology acceptance model is employed to evaluate behavioural attitudes both before and after exposure to self-driving cars. The objective is to glean insights into the perceived concerns associated with AV. These initial investigations set the groundwork for carrying out experiments on detecting pedestrians and cyclists using deep-learning based object detection algorithms analysing images captured from AVs. This thesis rigorously compared various deep-learning-based object detection algorithms using benchmark datasets (Cityscapes, Eurocity Person, and Kitti) to ensure efficiency and real-time safety for AV development. Five distinct algorithms (Faster RCNN, Cascade RCNN, FCOS, Deformable, DETR, and RetinaNet) were compared for the detection of urban road objects across various datasets. Additionally a diverse traffic benchmark dataset was constructed combining several benchmark datasets incorporating various weather, lighting, and traffic scenarios, evaluating state-of-the-art detection algorithms in traffic situations and adapting them accordingly. Lastly, the research made a significant contribution by developing a unified algorithm for simultaneous estimation, tracking, and detection of multiple road objects, advancing AV technology and promoting safer autonomous transportation systems. This algorithm was tested on real-life traffic images collected through afield study. This thesis underscores the paramount importance of comprehending public perception and acceptance of autonomous vehicles (AVs). Simultaneously, it emphasizes the necessity for the development of robust and efficient detection algorithms to ensure the safety of AVs in diverse, real-world environments. The creation of a comprehensive traffic benchmark dataset is a notable contribution, addressing the need for varied and challenging test cases in the advancement of autonomous vehicle technology. Furthermore, the establishment of a unified algorithm for multi-object estimation and tracking fills a crucial research gap, providing a foundational element for holistic approaches to AV technology. Collectively, this research marks a substantial leap forward in the continual progress of autonomous vehicles, offering valuable insights into societal acceptance, safety enhancement through advanced algorithms, and the provision of essential tools for the ongoing development and integration of AVs into contemporary transportation systems