Lane-Guided Monocular 3D Object Detection for Autonomous Driving

Abstract

Contemporary navigation systems employed in robotics and autonomous vehicles heavily rely on 3D data processing technologies, underscoring the significance of accurate recognition and analysis of the surrounding environment. Estimating the 3D position and orientation of a target object within its surroundings using monocular RGB cameras constitutes a crucial yet challenging task for autonomous navigation systems. Consequently, ongoing research endeavors are actively exploring approaches that leverage the information available in monocular RGB images or incorporate additional data, such as generating pseudo-LiDAR values.

Given that autonomous driving scenarios predominantly occur on roadways, lanes assume a paramount role as a crucial input. Thus, it is posited that the analysis of lane information can enhance object detection in the surrounding environment and further bolster the system's situational understanding capabilities.

Within the context of this thesis, a novel network is proposed that harnesses lane information to effectively exploit line-based cues derived from monocular RGB images. Existing studies have employed deep Hough transform networks to extract lines from images using line information from monocular RGB images. However, the utilization of deep Hough transform networks entails limitations, such as extracting unnecessary lines in autonomous driving scenarios or inadequately capturing important lines. To overcome these challenges, the proposed approach incorporates lane information extracted through the learning of a lane detection network. Consequently, only essential and meaningful lane information is learned, eschewing unnecessary or inaccurate line information encountered with existing deep Hough transforms.

The incorporation of lane information learning enables a more precise estimation of the depth of 3D objects. Experimental evaluations conducted on the KITTI 3D Object Detection dataset demonstrate that the technique proposed in this thesis surpasses existing methods, substantiating its efficacy and potential advancements in the field.