Visual Background Extractor Using Randomized History Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing background detection methods struggle with adapting to a wide range of background evolution rates and are vulnerable to noise, requiring complex parameter tuning and additional post-processing for spatial coherence, while also being inefficient in initialization and computation-intensive.
Innovation Solution
A background detection method that uses a random selection of sample pixel values for updating and initialization, incorporating auxiliary histories for different time responses, and randomizing the update process to reduce bias and computation, allowing for sharper segmentation and noise resilience without complex parameter estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deterministic background detection methods are used, then spatial coherence can be achieved through post-processing, but the computational complexity and parameter tuning requirements increase
Solution Approach 1:
The patent applies self-service by using self-organizing maps that automatically learn and adapt to background patterns without requiring manual parameter tuning. The neural network structure performs unsupervised learning, allowing the system to autonomously detect background evolution rates and adjust to changes, eliminating the need for complex post-processing and parameter adjustment while maintaining spatial coherence
Solution Approach 2:
The patent implements dynamics by using adaptive background models that can adjust to varying background evolution rates. The system dynamically adapts to different rates of background change through the neural network's learning mechanism, allowing it to handle both slow and fast background variations without requiring deterministic post-processing steps
2Measurement precision
If complex parameter estimation and post-processing are used, then detection accuracy improves, but computational efficiency decreases
Solution Approach 1:
The patent replaces traditional mechanical post-processing operations with a neural network-based system. Instead of using deterministic algorithms that require multiple processing passes and parameter adjustments, the system uses self-organizing maps that perform background detection in a single pass, substituting complex mechanical computation with adaptive neural processing that achieves both accuracy and efficiency
Solution Approach 2:
The patent applies preliminary action by pre-training the self-organizing map with background data before actual detection. This preliminary learning phase allows the system to establish background patterns in advance, so that during runtime, background detection can be performed rapidly without requiring complex real-time parameter estimation or post-processing operations
3Measurement precision
If extensive initialization is performed, then background model accuracy improves, but initialization time increases
Solution Approach 1:
The patent applies partial action by using a streamlined initialization process that only performs essential background learning without exhaustive parameter tuning. The self-organizing map is initialized with a subset of necessary background data, allowing the system to achieve adequate accuracy quickly, and then continues to learn and refine background patterns during normal operation rather than requiring complete initialization beforehand
4Stability of the object's composition
If deterministic update methods are used, then background consistency is maintained, but adaptability to varying background evolution rates decreases
Solution Approach 1:
The patent implements dynamics by using adaptive background models that can adjust to varying background evolution rates. The system dynamically adapts to different rates of background change through the neural network's learning mechanism, allowing it to handle both slow and fast background variations while maintaining consistency through the self-organizing structure that preserves spatial relationships
Data Source
AI summary
The present invention relates to a Visual Background Extractor (VIBE) consisting in a method for detecting a background in an image selected from a plurality of related images. Each one of said set of images is formed by a set of pixels, and captured by an imaging device. This background detection method comprising the steps of: establishing, for a determined pixel position in said plurality of images, a background history comprising a plurality of addresses, in such a manner as to have a sample pixel value stored in each address; comparing the pixel value corresponding to said determined pixel position in the selected image with said background history, and, if said pixel value from the selected image substantially matches at least a predetermined number of said sample pixel values: classifying said determined pixel position as belonging to the image background; and—updating said background history by replacing the sample pixel values in one randomly chosen address of said background history with said pixel value from the selected image. The method of the invention is applicable a.o. for video surveillance purposes, videogame interaction and imaging devices with embedded data processors.