Object Detection Based on the Fusion of Sparse LiDAR Point Cloud and Dense Stereo Pseudo Point Cloud

3D object detection is critical for unmanned intelligent vehicles to perceive their environment. LiDAR and stereo cameras are commonly used sensors, but each has limitations. LiDAR provides high-precision 3D point cloud data, but sparse distant and small objects pose challenges. Stereo cameras gener...

Full description

Saved in:
Bibliographic Details
Published in2024 4th International Conference on Neural Networks, Information and Communication (NNICE) pp. 860 - 863
Main Authors Li, Yanming, Su, Jianqiang, Liu, Liqiang, Liu, Peng
Format Conference Proceeding
LanguageEnglish
Published IEEE 19.01.2024
Subjects
GDC fusion algorithm
Intelligent vehicles
Laser radar
LiDAR
mask operation
Object detection
Pedestrians
Point cloud compression
stereo camera
Technological innovation
Three-dimensional displays
threshold filtering
Online AccessGet full text

Cover

More Information
Summary:3D object detection is critical for unmanned intelligent vehicles to perceive their environment. LiDAR and stereo cameras are commonly used sensors, but each has limitations. LiDAR provides high-precision 3D point cloud data, but sparse distant and small objects pose challenges. Stereo cameras generate dense pseudo point clouds, but lack accurate positioning. This article combines the strengths of dense stereo cameras and precise LiDAR to improve detection performance for distant and small objects. We propose two innovations: (1) a method that combines 2D object detection and threshold filtering to accurately identify small distant objects, and (2) a local Graph based Depth Correction (GDC) that introduces a mask operation to correct and fusion small and distant object areas, filter out invalid pseudo point clouds, and improve accuracy. Experimental results on the PV-RCNN network demonstrate improved detection accuracy for distant small objects, with increased performance levels across Easy, Moderate, and Hard categories for pedestrian small objects (15.88%, 16.86%, and 14.41% respectively).
DOI:10.1109/NNICE61279.2024.10498214