3D Point Cloud Bounding Boxes for Occluded Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional point cloud object detection methods face challenges in accurately determining 3D center point coordinates due to objects being covered, leading to reduced accuracy in 3D bounding box detection, which affects the overall object detection precision.
Innovation Solution
A method that involves obtaining 3D object bounding boxes using 3D coordinates of bounding box corner points by detecting minimum and maximum values on the Z-axis, splitting the 3D bounding box into two 2D bounding boxes, and using a preset point cloud object detection model to enhance accuracy, with visibility-based loss weight adjustments and orientation correction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional CSA mode is used to obtain 3D bounding box, then the detection process is simple, but the accuracy deteriorates when objects are covered and 3D point clouds are missing
Solution Approach 1:
The patent segments the 3D bounding box detection into multiple 2D bounding box detections on different planes (front, back, left, right, top, bottom). Each 2D bounding box is detected independently, and their intersections determine the final 3D bounding box. This segmentation allows the system to handle occluded objects more robustly by relying on visible surfaces rather than requiring complete 3D point clouds.
Solution Approach 2:
The patent transforms the 3D bounding box detection problem into multiple 2D bounding box detection problems on different planes. By detecting 2D bounding boxes on front, back, left, right, top, and bottom planes and computing their intersections, the system achieves more accurate 3D bounding box estimation, particularly when objects are partially occluded and 3D point clouds are incomplete.
2Measurement precision
If 3D center point coordinates are used to represent bounding box, then the representation is compact, but the accuracy deteriorates when objects are covered
Solution Approach 1:
The patent divides the single 3D center point estimation into multiple 2D center point estimations from different viewing planes. Each 2D plane provides independent center point information, and the combination of these plane-specific centers yields a more accurate overall center point location, even when parts of the object are occluded.
Solution Approach 2:
The patent applies different detection strategies to different parts of the object by detecting 2D bounding boxes on multiple planes with different orientations. Each plane's detection is optimized for its local viewing conditions, and the combination of these local detections provides globally accurate center point coordinates even when local 3D point clouds are missing.
Data Source
AI summary
The disclosure relates to the technical field of autonomous driving, and specifically provides a point cloud object detection method, a computer device, a storage medium, and a vehicle, to solve the problem of improving the accuracy of point cloud object detection. The method includes: obtaining a three-dimensional (3D) point cloud frame collected by a radar, performing object detection on the 3D point cloud frame to obtain a 3D object bounding box represented by 3D coordinates of bounding box corner points, and obtaining an object detection result based on the 3D object bounding box. Through the method, even if an object is covered, coordinates of uncovered end points of the object can be accurately obtained based on 3D coordinates of bounding box corner points in a 3D object bounding box, so that the accuracy of object detection can be effectively improved, and effective tracking corner points are provided for object tracking, thereby ensuring the accuracy and reliability of object tracking.