CNN for Detecting Objects in Autonomous Vehicle Environments using 2D Images: Current Solutions, Challenges and Future Trends

Abstract

Object recognition within the surrounding environment is pivotal for the advancement and safety of autonomous vehicles (AV). It is imperative for these vehicles to precisely identify and differentiate between key objects, such as pedestrians, traffic lights, lanes, and other vital elements. Any inaccuracy in this domain can have severe repercussions, potentially resulting in catastrophic accidents. Recent methodologies have capitalized on the capabilities of Convolutional Neural Networks (CNN) to enhance object detection within the AV domain. In the context of AV, data acquisition predominantly stems from two primary sources: (a) Light Detection and Ranging (LiDAR) systems, which produce a 3D Point Cloud representation of objects, and (b) cameras, offering a 2D visual perspective of the environment. Notably, there has been a predominant emphasis on leveraging 3D data in contemporary research. This paper seeks to address this research gap by concentrating on 2D object detection techniques in the AV environment. We will discuss the latest advancements in object detection for AVs harnessing CNNs, with a focus on Lane Detection, Pedestrian Recognition, and the detection of Traffic Signs and Lights. Furthermore, we will present an overview of emerging datasets created explicitly for AV research, detailing their unique specifications. The paper will conclude by shedding light on current challenges and forecasting prospective trends in AV object detection.