Pedestrian Retrieval Using Multidimensional Feature Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing pedestrian retrieval methods in video surveillance face low accuracy due to small similarities between pedestrian images caused by posture, light, and other factors, leading to ineffective clustering of multidimensional feature data.
Innovation Solution
A pedestrian retrieval method that extracts body, upper-body, and lower-body multidimensional feature data from target images, using locality-sensitive hash values and binary codes for screening stored feature data to filter out noise and improve retrieval efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional bag-of-words model with clustering is used for pedestrian retrieval, then the system can handle large quantities of pedestrian images, but the retrieval accuracy is low due to poor clustering effect caused by posture, light, and other factors
Solution Approach 1:
The patent segments the pedestrian image into multiple body parts (head, upper body, lower body) and extracts features independently for each part. This segmentation allows the system to capture local characteristics that are more discriminative than global features alone, thereby improving retrieval accuracy while maintaining manageable complexity through modular processing
Solution Approach 2:
The patent transforms the feature representation by creating a hierarchical structure with both local body part features and global whole-body features. This dimensional transformation from flat feature vectors to hierarchical feature sets enables more effective comparison and matching, resolving the accuracy-complexity contradiction
2Measurement precision
If multidimensional feature data is extracted from all pedestrian images for comprehensive matching, then retrieval accuracy improves, but the amount of data to be processed increases significantly
Solution Approach 1:
The patent extracts and stores features for specific body parts (head, upper body, lower body) separately from the complete pedestrian image. This selective extraction reduces the dimensionality of the feature space while preserving the most discriminative local characteristics, thereby maintaining accuracy with reduced data quantity
Solution Approach 2:
The patent uses a two-stage approach where first local body part features are matched to generate candidate results, then whole-body features are used for final verification. This partial action strategy processes only necessary features at each stage, reducing overall computational load while maintaining comprehensive matching accuracy
3Measurement precision
If strict matching criteria are applied to ensure high retrieval accuracy, then fewer false positives occur, but the retrieval speed decreases due to increased computational requirements
Solution Approach 1:
The patent implements a two-stage retrieval process with periodic filtering: first stage uses local body part features for rapid candidate generation, second stage applies stricter whole-body feature matching for final verification. This periodic application of different matching strictness levels maintains high accuracy while preserving retrieval speed through progressive filtering
Solution Approach 2:
The patent performs preliminary matching using local body part features before applying comprehensive whole-body feature matching. This preliminary action filters out obviously mismatched candidates early, reducing the computational burden of subsequent strict matching operations and maintaining both speed and accuracy
Data Source
AI summary
A pedestrian retrieval method and apparatus that belong to the video surveillance field include extracting first feature data, second feature data, and third feature data of a target pedestrian image, where the target pedestrian image is an image of a to-be-retrieved pedestrian, and the first feature data, the second feature data, and the third feature data respectively include a plurality of pieces of body multidimensional feature data, a plurality of pieces of upper-body multidimensional feature data, and a plurality of pieces of lower-body multidimensional feature data of the target pedestrian image, screening stored multidimensional feature data based on the first feature data, the second feature data, and the third feature data to obtain a target feature data set, and outputting a pedestrian retrieval result using the target feature data set.