Pedestrian Detection via GMM Feature Dimension Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition technologies face challenges in improving detection accuracy while reducing calculation costs, particularly when quantizing luminance gradients for feature extraction, as narrow quantization intervals lead to increased feature amounts and processing times.
Innovation Solution
The proposed solution involves generating a probability density function based on the occurrence frequency of co-occurrence in luminance gradient directions, using a Gaussian Mixture Model (GMM) to dynamically form histograms, and reducing the number of basis functions to decrease feature dimensions, thereby reducing calculation costs without compromising detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the interval of quantization is narrowed to improve accuracy, then detection accuracy is improved, but the feature amount becomes enormous and processing time increases
Solution Approach 1:
The patent changes the parameter representation from discrete quantized histogram bins to continuous probability density functions. Instead of using fixed quantization intervals that create enormous feature vectors when narrowed, the system uses PDFs with parameters (mean, variance, skewness, kurtosis) that can be computed efficiently regardless of the desired precision level, thus resolving the contradiction between accuracy and feature complexity
Solution Approach 2:
The patent replaces the mechanical quantization process (which creates discrete bins and suffers from the curse of dimensionality when intervals are narrowed) with a statistical approach using probability density functions. This substitution allows continuous representation of gradient direction distributions without the computational explosion associated with fine quantization, achieving both high accuracy and manageable complexity
2Measurement precision
If the interval of quantization is narrowed to improve accuracy, then detection accuracy is improved, but processing time increases
Solution Approach 1:
The patent transforms the computation from calculating high-dimensional histograms requiring iterative quantization to computing statistical moments (mean, variance, skewness, kurtosis) of gradient directions. These moment calculations can be performed efficiently using standard algorithms without the computational burden of fine quantization, thereby reducing processing time while maintaining or improving detection accuracy
Solution Approach 2:
The patent replaces the time-consuming quantization and histogram accumulation process with a statistical moment calculation approach. By computing PDF parameters directly from gradient data without creating discrete bins, the system avoids the computational overhead associated with fine quantization intervals, thus reducing processing time while preserving detection accuracy
Data Source
AI summary
An image processing device has a function for plotting a luminance gradient co-occurrence pair of an image on a feature plane and applying an EM algorithm to form a GMM. The device learns a pedestrian image and creates a GMM, subsequently learns a background image and creates a GMM, and calculates a difference between the two and generates a GMM for relearning based on the calculation. The device plots a sample that conforms to the GMM for relearning on the feature plane by applying an inverse function theorem. The device forms a GMM that represents the distribution of samples at a designated mixed number and thereby forms a standard GMM that serves as a standard for image recognition. When this mixed number is set to less than a mixed number designated earlier, the dimensions with which an image is analyzed are reduced, making it possible to reduce calculation costs.