3D Obstacle Detection Accuracy via 2D Projection Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The accuracy of 3D obstacle detection in intelligent transportation systems is compromised due to variations in deployment locations, angles, and camera intrinsic parameters, leading to poor prediction of obstacle positions using existing 3D detection technologies.
Innovation Solution
The method optimizes 3D obstacle detection by utilizing position information from both 2D and 3D detection frames, along with 2D projection frames, through a loss function constructed using constraint items based on central points, sides, and projection points, to improve the accuracy of 3D position prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If 3D detection technology is applied to detect obstacle position, then the system can obtain three-dimensional position information, but the accuracy of 3D position prediction deteriorates due to variations in deployment locations, angles and camera intrinsic parameters
Solution Approach 1:
The patent employs a feedback mechanism by using the 2D detection results as correction feedback to the 3D detection results. The loss function calculates the difference between projected 3D bounding boxes and actual 2D bounding boxes, and this error feedback is used to optimize the 3D position parameters, thereby improving accuracy while adapting to various deployment conditions
Solution Approach 2:
The patent changes the parameter optimization approach by introducing a loss function that adjusts 3D position parameters based on 2D projection constraints. Instead of relying solely on camera calibration parameters that vary with deployment location and angle, the system optimizes 3D parameters by enforcing geometric consistency with 2D observations, making the system more adaptable to different deployment scenarios
2Measurement precision
If camera intrinsic parameters are used for 3D detection, then the system can transform 2D images to 3D space, but the accuracy deteriorates due to parameter variations across different deployment locations
Solution Approach 1:
The patent introduces 2D projection frames as an intermediary between 3D detection results and actual 2D observations. This intermediary layer projects 3D bounding boxes into 2D image space and compares them with actual 2D bounding boxes, serving as a mediator that eliminates the need for complex camera intrinsic parameter transformations while maintaining geometric accuracy
Solution Approach 2:
The patent creates a virtual copy of the 3D detection result by projecting it into 2D space to form a 2D projection frame. This copied 2D representation is then used for comparison and optimization, allowing the system to work with simplified 2D geometry instead of complex 3D camera calibration parameters, thereby reducing device complexity while maintaining precision
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
An obstacle detection method and device, an apparatus, and a storage medium are provided, which are related to a field of intelligent transportation. The specific implementation includes: acquiring position information of a two-dimensional (2D) detection frame and position information of a three-dimensional (3D) detection frame of an obstacle in an image; converting the position information of the 3D detection frame of the obstacle into position information of a 2D projection frame of the obstacle; and optimizing the position information of the 3D detection frame of the obstacle by using the position information of the 2D detection frame, the position information of the 3D detection frame and the position information of the 2D projection frame of the obstacle in the image. Accuracy of results of predicting a 3D position of an obstacle by a roadside, on-board sensing device, or other sensing devices may be improved.