Adaptive Feature Extraction for Thermal Pedestrian Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional pedestrian detection in thermal images faces challenges due to the curse of dimensionality and feature correlation, leading to increased detection errors and reduced computational efficiency.
Innovation Solution
An image adaptive feature extraction method that divides images into blocks, applies feature extraction using SVM classifiers to obtain hyperplane normal vectors, sets thresholds, and integrates adaptive feature blocks to form adaptive feature images, reducing dimensionality and error probability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If feature vectors are directly cascaded to increase distinguishability, then detection accuracy is improved, but the dimensionality increases causing the curse of dimensionality problem and reducing computational efficiency
Solution Approach 1:
The patent extracts and selects only the most discriminative features from the cascaded feature vectors using SVM-based feature selection. This extracts the essential information needed for accurate detection while removing redundant dimensions, thereby maintaining detection accuracy while reducing the curse of dimensionality
Solution Approach 2:
The patent changes the parameter representation by transforming the high-dimensional cascaded features into a lower-dimensional space through SVM hyperplane projections. This parameter transformation preserves the discriminative power while reducing dimensionality, solving the curse of dimensionality problem
2Measurement precision
If multiple features of different characteristics are integrated to improve detection accuracy, then distinguishability is enhanced, but feature correlation causes overlapping data and increases detection errors
Solution Approach 1:
The patent applies SVM-based parameter transformation that projects correlated features onto a new feature space defined by hyperplane normal vectors. This transformation decorrelates the features while preserving their discriminative power, eliminating overlapping data and reducing detection errors
Solution Approach 2:
The patent introduces SVM hyperplane normal vectors as intermediary representations that mediate between the original correlated features and the final detection decision. These normal vectors serve as a transformed feature space that eliminates correlation while maintaining discriminative information
3Reliability
If comprehensive feature extraction is performed on all image blocks to improve detection coverage, then detection completeness is improved, but computational efficiency is reduced
Solution Approach 1:
The patent extracts and processes only the most relevant image blocks that contain potential pedestrian features, identified through SVM-based relevance assessment. This selective extraction maintains detection completeness by focusing on informative regions while significantly reducing computational load on irrelevant blocks
Data Source
AI summary
An image adaptive feature extraction method includes dividing an image into a plurality of blocks, performing a feature extraction processing on the plurality of blocks, and obtaining a block feature from each of the plurality of blocks after the feature extraction processing; calculating each block feature by means of a support vector machine (SVM) classifier, wherein each block feature is calculated to obtain a hyperplane normal vector; setting a threshold value, determining the block feature according to the hyperplane normal vector, recording the block as an adaptive feature block when a value of the hyperplane normal vector is higher than the threshold value, and integrating each adaptive feature block to form an adaptive feature image. Because an image adaptive feature extraction process is performed before a pedestrian image detection is calculated, and effective feature data is then selected, computational efficiency is boosted and detection pedestrian error probability is reduced.