Voting Space Loop Closure for AR Drift Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Augmented reality devices experience positional drift during visual odometry, leading to incorrect display of virtual content due to cumulative errors, especially in scenes with varying viewpoints, textures, and reflective surfaces, which conventional methods using image descriptors fail to address effectively.
Innovation Solution
The system employs an accumulator voting space to limit the possible camera locations, relying on the rotational part of the pose, camera height, and 3D coordinates of tracked points to correct the estimated pose, avoiding the use of image descriptors and thus reducing drift by linking the current image to previously reconstructed 3D map points.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional methods using image descriptors are used for loop closure detection, then the system can handle general scenes, but the measurement precision deteriorates due to cumulative errors and failure in scenes with varying viewpoints, textures, and reflective surfaces
Solution Approach 1:
The patent extracts and removes image descriptors from the loop closure detection process. Instead of using traditional image descriptor matching, the system directly compares 3D map points between current and historical images, eliminating the source of cumulative errors that plague conventional descriptor-based methods.
Solution Approach 2:
The patent replaces the mechanical/image-processing system of descriptor extraction and matching with a direct 3D spatial coordinate comparison system. This substitution transitions from 2D image space operations to 3D world space operations, fundamentally changing how loop closure is detected and eliminating drift accumulation.
2Measurement precision
If the system uses 3D map points for loop closure detection, then the measurement precision improves by reducing drift, but the device complexity increases due to the need for maintaining and querying 3D maps
Solution Approach 1:
The patent performs preliminary actions by pre-constructing and storing 3D maps during the initial exploration phase. These 3D maps, containing spatial coordinates of map points, are prepared in advance so that during loop closure detection, the system only needs to query and compare coordinates rather than perform complex image processing, thus reducing real-time computational complexity.
Solution Approach 2:
The patent creates a 3D copy of the environment through map construction, storing spatial information in a simplified data structure. This 3D map copy allows for efficient querying and comparison operations during loop closure detection, reducing the complexity of real-time processing while maintaining high positional accuracy.
3Productivity
If the system limits camera location possibilities using voting space, then the productivity improves by reducing search space, but the measurement precision may worsen due to potential false positives in the voting process
Solution Approach 1:
The patent segments the loop closure detection process into distinct phases: 3D map point identification, voting space construction, and peak detection. By dividing the problem into these manageable segments, the system can efficiently process each phase independently, improving overall productivity while maintaining precision through the structured approach.
Solution Approach 2:
The voting space mechanism provides feedback by accumulating votes from multiple 3D map point comparisons. The system uses this feedback to identify dominant peaks that indicate loop closure, with the voting process itself serving as a verification mechanism that reduces false positives while maintaining high processing speed.
Data Source
AI summary
A method for detecting a loop closure is described. A device accesses pose information and a three-dimensional map of feature points generated by a visual inertia system of the device. The device splits the pose information into a translational part and a rotational part. The device limits the translational part to two-dimensional coordinates and estimates two-dimensional information of the limited translational part based on an accumulator voting space. The device determines an updated pose of the device based on the estimated two-dimensional information, the rotational part, and the three-dimensional map. The pose information is updated with the updated pose.