3D Bounding Box Projection for Accurate Vehicle Anchor Points
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for object recognition and position estimation in 3D from 2D images are prone to errors and require complex assumptions or calibration data, especially when dealing with vehicles in traffic scenarios, leading to inaccurate lane allocation, traffic counting, and collision detection.
Innovation Solution
A method using a neural network to determine a 2D frame projection of a 3D bounding box directly in a camera image without relying on 3D world coordinates or calibration data, allowing for robust estimation of corner and anchor points for objects like vehicles, which are then used for precise monitoring tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard object detection methods are used in 2D camera images, then object recognition can be performed, but anchor point estimation and position determination become inaccurate due to large camera aperture angles and special camera poses
Solution Approach 1:
The patent transitions from 2D image-based anchor point estimation to 3D space-based estimation by projecting 3D bounding boxes onto the 2D image plane. This dimensional transformation allows the system to leverage three-dimensional spatial relationships and camera pose information to accurately determine anchor points, thereby resolving the inaccuracy caused by large aperture angles and special camera poses.
Solution Approach 2:
The patent introduces a 3D bounding box projection as an intermediary between the 2D image and the 3D world coordinates. This intermediary mechanism enables the system to bridge the gap between 2D image observations and 3D spatial understanding, allowing for accurate anchor point estimation even under challenging camera configurations.
2Measurement precision
If 3D considerations are used for object detection, then position estimation improves, but processing complexity and computational requirements increase
Solution Approach 1:
The patent performs preliminary actions by pre-defining the 3D bounding box geometry and projection relationships before actual object detection. This preparation allows the system to efficiently compute 3D positions from 2D images using standardized transformation matrices and pre-calculated camera parameters, reducing real-time computational complexity while maintaining 3D position estimation precision.
Solution Approach 2:
The patent creates a virtual copy of the 3D bounding box by projecting it onto the 2D image plane. This copying approach allows the system to work with simplified 2D representations while retaining the benefits of 3D spatial understanding, thereby reducing processing complexity compared to full 3D reconstruction methods.
3Area of stationary object
If camera aperture angle is increased to capture more scene, then field of view improves, but anchor point estimation accuracy deteriorates
Solution Approach 1:
The patent changes the fundamental parameters of position estimation from 2D image coordinates to 3D space coordinates by incorporating camera pose information. This parameter transformation allows the system to maintain accurate anchor point estimation even when using large aperture angles and wide field of view, as the 3D projection model compensates for the geometric distortions introduced by wide-angle cameras.
Data Source
AI summary
In a method (30) for obtaining information (18a, b) about a potential object (8a, b) in a camera image (10) of a scene (2), A) the camera image (10) is provided as a single 2D camera image, B) potential objects (8a, b) are detected in the camera image (10), C) a frame projection (14a, b) of a 3D frame (16) directly surrounding the object (8a, b) into the camera image (10) is determined for at least one of the detected objects (8a, b) as at least part of the information (18a, b), wherein steps B) and C) are each performed with the aid of a neural network (32).