Stochastic Ranging for Fast Dense Depth Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computational stereo techniques face challenges in determining robust correspondences between pixels in images, especially in uncontrolled outdoor scenes with moving objects and varying lighting, leading to inefficiencies and inaccuracies in dense reconstruction.
Innovation Solution
A stochastic method for fast stereoscopic ranging is introduced, which selects a pair of images, seeds estimated values for a range metric, computes local influences, aggregates these influences, and refines the estimates through iterative post-processing, utilizing a combination of cooperative and stochastic techniques to reduce computation and improve robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional computational stereo techniques are used to determine robust correspondences between pixels, then depth estimation can be performed, but the computation time increases and accuracy decreases in uncontrolled outdoor scenes with moving objects and varying lighting
Solution Approach 1:
The patent divides the image into superpixels or patches and processes them in groups rather than individual pixels. This segmentation reduces the total number of correspondence checks while maintaining depth estimation accuracy, thereby reducing computation time in complex outdoor scenes
Solution Approach 2:
The patent performs preliminary processing by identifying and masking out moving objects and varying lighting regions before depth estimation. This preliminary action removes problematic elements that would otherwise require extensive computation, reducing overall computation time while maintaining accuracy in stable regions
2Reliability
If dense reconstruction is performed on all pixels in the imagery, then complete depth coverage is achieved, but the computational complexity and processing time increase significantly
Solution Approach 1:
The patent segments the image into superpixels or patches and performs dense reconstruction on these grouped elements rather than individual pixels. This maintains complete depth coverage while reducing computational complexity by processing fewer, larger units
Solution Approach 2:
The patent applies dense reconstruction selectively to certain regions or uses a reduced set of processing steps for areas where full density is less critical, achieving sufficient depth coverage without the full computational burden of processing every pixel with equal detail
3Measurement precision
If correspondence matching is performed exhaustively to ensure accuracy, then robust depth estimates are obtained, but the processing speed decreases
Solution Approach 1:
The patent groups pixels into superpixels or patches and performs correspondence matching on these groups. This reduces the number of individual matches required while maintaining accuracy through the collective information from multiple pixels, thereby increasing processing speed
Solution Approach 2:
The patent combines information from multiple pixels within superpixels or patches to form correspondence matches. This merging approach maintains robustness by using aggregated evidence from multiple pixels while reducing the total number of matching operations required, thus improving processing speed
Data Source
AI summary
A stochastic method and system for fast stereoscopic ranging includes selecting a pair of images for stereo processing, in which the pair of images are a frame pair and one of the image is a reference frame, seeding estimated values for a range metric at each pixel of the reference frame, initializing one or more search stage constraints, stochastically computing local influence for each valid pixel in the reference frame, aggregating local influences for each valid pixel in the reference frame, refining the estimated values for the range metric at each valid pixel in the reference frame based on the aggregated local influence, and post-processing range metric data. A valid pixel is a pixel in the reference frame that has a corresponding pixel in the other frame of the frame pair. The method repeats n iterations of the stochastically computing through the post-processing.