ToF Range Image Denoising via Per-Pixel Calibration and Geodesic Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Time-of-flight (ToF) cameras suffer from both scene-independent and scene-dependent noise, which limit their accuracy in range image measurements due to manufacturing limitations and multipath interference, respectively, and existing methods fail to effectively model pixel location-dependent biases and correct measurement distortions.
Innovation Solution
A method using a per-pixel calibration model with a planar checkerboard pattern and feedforward neural networks to reduce scene-independent noise, and a geodesic filter based on confidence values and edge locations to address scene-dependent noise, leveraging a dataset of ToF and ground truth range images to train neural networks for denoising.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a global calibration model is used to reduce scene independent noise, then the model complexity is low and requires small data for fitting, but the measurement precision deteriorates because pixel location dependent bias cannot be modeled
Solution Approach 1:
The patent divides the calibration process into per-pixel calibration units, where each pixel is calibrated independently using individual calibration parameters. This segmentation allows the system to capture pixel location dependent biases that a global model would miss, thereby improving measurement precision while keeping each individual pixel's calibration simple.
Solution Approach 2:
The patent applies local quality by using per-pixel calibration parameters that are specific to each pixel's location and characteristics. This allows different parts of the image sensor to be calibrated according to their local properties, capturing spatial variations in bias that a global model cannot address.
2Measurement precision
If simulation-based methods with multiple modulation frequencies are used to reduce scene dependent noise, then the measurement precision improves, but the productivity deteriorates due to slow processing speed and special hardware requirements
Solution Approach 1:
The patent changes the approach from using multiple modulation frequencies to using a single frequency with a learned filtering model. The neural network is trained on datasets captured at different frequencies and conditions, then applies learned parameters to single-frequency data in real-time, achieving high precision without the computational burden of multi-frequency processing.
Solution Approach 2:
The patent creates a learned model (copy of the denoising function) through neural network training that replicates the效果 of complex multi-frequency processing. This learned model can then be applied rapidly to single-frequency data, achieving both high precision and fast processing speed.
3Object-affected harmful factors
If conventional image denoising methods are used on ToF range images, then random noises are removed, but the measurement precision deteriorates because measurement biases and fine structures cannot be corrected
Solution Approach 1:
The patent introduces an intermediary step of per-pixel calibration that corrects measurement biases before applying denoising. The calibration parameters act as intermediaries that adjust each pixel's measurements based on its location-dependent characteristics, preserving fine structures while removing random noise in subsequent processing steps.
Data Source
AI summary
A method for denoising a range image acquired by a time-of-flight (ToF) camera by first determining locations of edges, and a confidence value of each pixel, and based on the locations of the edges, determining geodesic distances of neighboring pixels. Based on the confidence values, reliabilities of the neighboring pixels are determined and scene dependent noise is reduced using a filter.