Pose Estimation Using Visual and Inertial Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing pose estimation methods in autonomous driving face challenges with accuracy and stability due to GPS signal hopping in occluded environments and error propagation from previous frame estimations, leading to drift in camera pose estimation results.
Innovation Solution
A pose estimation method that combines a visual observation model constraint and a motion model constraint using multi-frame images and inertial data from an inertial measurement unit, allowing for accurate and robust pose estimation by decoupling state variables and improving real-time performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fusion optimization method based on camera images and GPS data is used, then position and posture information can be obtained, but accuracy and stability deteriorate due to GPS signal hopping in occluded environments
Solution Approach 1:
The patent combines visual observation model constraints from camera images with motion model constraints from inertial measurement unit data to form a fused constraint system. This merging allows the system to leverage the strengths of both sensors while compensating for their individual weaknesses, particularly using IMU data to maintain stability when GPS signals are unavailable or unreliable in occluded environments.
Solution Approach 2:
The inertial measurement unit acts as an intermediary between GPS and visual systems. When GPS signals are blocked or visual features are insufficient, the IMU provides continuous motion constraints that bridge the gap, preventing pose estimation drift and maintaining system reliability during transitions between different sensing conditions.
2Productivity
If camera pose estimation of consecutive frames is based on previous frame estimation results, then continuous pose information can be obtained, but pose errors are propagated and accumulated leading to increased drift
Solution Approach 1:
The patent implements a feedback mechanism where the visual observation model constraints from each frame are used to correct and reset the cumulative pose estimation. By continuously comparing observed visual features with predicted features based on motion model propagation, the system detects and corrects drift errors, preventing error accumulation while maintaining real-time estimation capability.
Solution Approach 2:
The motion model constraints are established in advance based on inertial measurement data before visual processing occurs. This preliminary constraint framework provides a predicted pose trajectory that guides the visual feature matching process, reducing the propagation of errors by constraining the search space and preventing divergent error accumulation.
3Loss of information
If GPS-based fusion optimization is used for pose estimation, then position information can be obtained, but signal hopping occurs in occluded scenes such as tunnels and tall buildings
Solution Approach 1:
The patent applies different constraint qualities to different spatial and temporal contexts. In open environments with GPS availability, the system uses GPS-based fusion optimization. In occluded environments like tunnels or under tall buildings, the system automatically transitions to relying on visual observation constraints and motion model constraints from the IMU, ensuring continuous and reliable pose estimation without signal hopping.
Data Source
AI summary
The present application discloses a pose estimation method and device, related equipment and a storage medium. In the pose estimation method, a visual observation model constraint is constructed according to images, a motion model constraint is constructed according to inertial data, and a pose of a camera, an inertial measurement unit or an object that corresponds to the time of each image is calculated according to the visual observation model constraint and the motion model constraint. The pose estimation method adopts a fusion model of the visual observation model constraint and the motion model constraint to improve the accuracy, stability and robustness of pose estimation in a complex scene.