Obstacle Pose Annotation via Re-Projection Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic annotation methods for obstacles in autonomous driving face limitations due to sparse point clouds, poor intrinsic and external parameter calibration, and inconsistent image-point cloud synchronization, leading to limited annotation accuracy and sight distance, especially for dynamic and static targets at long distances.
Innovation Solution
A method that optimizes a projection relationship based on re-projection error to establish a consistent obstacle coordinate system, allowing for accurate pose determination of obstacles in different frames without requiring intrinsic and external parameter calibration, using a bird's-eye view from multiple cameras to enhance annotation sight distance and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional automatic annotation methods are used, then the process is simple, but annotation accuracy and sight distance are limited due to sparse point clouds and poor calibration
Solution Approach 1:
The patent implements a feedback mechanism by calculating re-projection error between the projected obstacle position in the current frame and the actual detected position. This error feedback is used to iteratively optimize the projection relationship parameters, continuously improving annotation accuracy without requiring complex manual calibration processes.
Solution Approach 2:
The patent optimizes projection relationship parameters (intrinsic and external parameters) dynamically based on re-projection error minimization. By changing these parameters iteratively, the system achieves high annotation accuracy for both static and dynamic obstacles while maintaining automated operation, resolving the contradiction between precision and complexity.
2Measurement precision
If point cloud data is used for obstacle detection, then dynamic information can be captured, but annotation accuracy decreases at long distances due to sparsity
Solution Approach 1:
The patent introduces an intermediary approach by combining point cloud data with image data through a unified projection relationship. The image data serves as a supplement to the sparse point cloud information, providing additional constraints that enable accurate annotation of obstacles at long distances where point clouds become sparse.
Solution Approach 2:
The patent creates a universal annotation system that works for both static and dynamic obstacles using the same projection relationship optimization framework. The system universally handles different obstacle types by optimizing the projection parameters based on re-projection error, achieving consistent accuracy across various scenarios without requiring separate processing pipelines.
3Measurement precision
If intrinsic and external parameter calibration is performed, then projection accuracy improves, but computational burden and time increase
Solution Approach 1:
The patent implements self-service calibration where the system automatically optimizes its own projection parameters using re-projection error minimization. Instead of requiring external manual calibration, the system uses its detected obstacle positions and projected positions to self-correct and optimize the projection relationship, reducing both time and computational complexity while maintaining high accuracy.
Solution Approach 2:
The patent applies partial optimization by focusing the calibration process on optimizing the projection relationship parameters rather than performing complete re-calibration of all camera parameters. This partial action approach achieves sufficient projection accuracy for annotation purposes while significantly reducing computational time and complexity.
4Length of stationary object
If image data is used for obstacle detection, then annotation sight distance is extended, but consistency with point cloud data becomes difficult to maintain
Solution Approach 1:
The patent segments the annotation process into two independent parts: point cloud-based detection for dynamic obstacles and image-based detection for static obstacles. Each part uses its own optimized projection relationship, and the results are combined consistently through a unified coordinate transformation framework, maintaining data consistency while extending sight distance for both obstacle types.
Solution Approach 2:
The patent uses a unified projection relationship optimization as an intermediary that bridges point cloud data and image data. This intermediary optimization process ensures that both data sources are consistently transformed into the same coordinate system, maintaining data consistency while enabling extended annotation sight distance through image data utilization.
Data Source
AI summary
Provided is a method for automatically annotating an obstacle, an electronic device and a storage medium, relating to the field of artificial intelligence technology, and in particular, to technologies fields of autonomous driving, neural network, deep learning and the like. The method includes: optimizing a target parameter in a projection relationship based on a re-projection error, the projection relationship is used to project a target obstacle from a reference frame onto a frame to be optimized, and satisfies a constraint in which positions of the target obstacle in different frames are consistent in an obstacle coordinate system established according to the target obstacle; and determining a target pose of the target obstacle based on the optimized target parameter.