Comparative Study of Vision-Based and LiDAR-Based Object Detection Algorithms for Autonomous Driving

Abstract

Autonomous vehicles achieve motion perception and path planning through object recognition. With the integration of deep learning into object detection, numerous practical and innovative studies have emerged in this field. Considering both hardware costs and recognition accuracy, object detection algorithms can be categorized into vision-based and Light Detection and Ranging (LiDAR)-based approaches. This paper reviews classic literature and cutting-edge technologies of both methods, detailing how vision-based approaches utilize computer vision to implement the You Only Look Once (YOLO) series and Faster Region-Convolutional Neural Network (Faster-RCNN). It also explains LiDAR-based approaches, focusing on classic models of point cloud and image fusion, including early fusion, deep fusion, and late fusion architectures. The advantages and limitations of both methods are compared, with experiments on the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) dataset showing that vision-based single-stage detection algorithms perform excellently, while the LiDAR-based Cascade Bi-directional Fusion algorithm stands out. The paper concludes with promising research directions, highlighting that continuous hardware advancements and constant updates in object detection algorithms are bringing us closer to fully realizing the goal of autonomous driving.