CFPC: The Curbed Fake Point Collector to Pseudo-LiDAR-Based 3D Object Detection for Autonomous Vehicles

Engineering notes

It is superior to the baseline in most situations, particularly in the categories of cars and riders.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

3D object detection in autonomous driving systems perceives the surrounding environment and is the foundation for autonomous driving. Due to the sparsity inherent in point clouds in autonomous driving scenarios, LiDAR-based 3D object detection often fails to distinguish distant objects effectively. Addressing the issue of point cloud sparsity will enhance the detection range in autonomous driving scenarios. Pseudo point clouds have been used to enhance the ability of deep learning models to detect distant points. However, this approach has several shortcomings. In this paper, a curbed fake point collector (CFPC), which addresses the three issues caused by pseudo points, is proposed to support 3D object detection for autonomous vehicles. First, for noise points with inaccurate coordinates, the dead pixel checker (DPC) calculates the depth map gradient using the Sobel operator. This approach enables the deep learning model to identify noise points. Second, because of the excessive quantity of points, sparse prioritized local sampling (SPLS) reduces the number of input point clouds to a lightweight level that can be accommodated by computing devices with limited memory. This is achieved through grid-based random sampling and real-point-prioritized farthest point sampling. This module effectively samples an appropriate pseudo point cloud based on the density of points in local space. Third, with respect to interference among channels, channel mask set abstraction (CMSA) isolates channels describing different information within the point cloud using GroupMLP, which is an MLP that separates channels into their respective groups. Group separation facilitates the extraction of features without mutual influence, allocating half of the output channels to color information and the other half to geometric information. The effectiveness of our approach is demonstrated by the results of experiments conducted on the KITTI dataset. It is superior to the baseline in most situations, particularly in the categories of cars and riders.

Links and sources

• Official / arXiv page

Comments