Object Recognition for Autonomous Vehicles from Combined Color and LiDAR Data.

In recent years, autonomous driving vehicles have garnered substantial attention in both the commercial and scientific domains. A key challenge faced by these vehicles is the accurate detection and recognition of objects within complex real-world road environments, essential for their real-time decision-making capabilities. While color imaging has traditionally provided rich information, the utilization of LiDAR scanners presents advantages such as high-quality data collection under varying lighting conditions and the provision of precise spatial information with an extensive range. By combining data from color cameras and LiDAR scanners, the potential for object detection in autonomous driving is expanded, opening up new avenues for advancement. This paper introduces a novel 3D object detector that leverages a bird's-eye view map generated from a LiDAR point cloud along with RGB images as input data. It employs focal loss and Euler angle regression techniques to enhance object detection performance. Through ablation experiments, the achieved improvements are evaluated. Experimental results demonstrate the framerate and performance of the proposed 3D object detector, surpassing 46 frames per second and achieving an average precision exceeding 90%. Additionally, a more compact version of the detector is introduced, processing the same input data three times faster while maintaining reasonably high accuracy. [ABSTRACT FROM AUTHOR]