Enhancing 3D Detection Accuracy in Autonomous Driving through Pseudo-LiDAR Augmentation and Downsampling

3D object detection is crucial for autonomous driving. Currently, most approaches depend on costly LiDAR sensors for acquiring precise depth information. Although pseudo-LiDAR techniques based on stereo cameras offer a more cost-effective alternative, their performance is still inferior to LiDAR-based methods. The objective of this paper is to improve detection accuracy using pseudo-LiDAR augmentation combined with downsampling. Specifically, in terms of pseudo-LiDAR augmentation, the GDC algorithm is employed to correct depth errors based on 4-beam LiDAR clouds, thereby improving pseudo-LiDAR quality. For pseudo-LiDAR downsampling, a 64-beam point cloud downsampling algorithm is utilized to improve the efficiency of PV-RCNN++ on dense pseudo-LiDAR clouds. Results from experiments on the KITTI 3D dataset indicate that our method substantially boosts detection accuracy across three object categories compared to traditional stereo detection algorithms. Notably, in the task of car detection, it surpasses the state-of-the-art stereo detection algorithm DSGN++ in terms of accuracy.