Person Re-Identification Driven Localization Refinement for Search Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current person search methods suffer from low accuracy due to the conflict between shared features in joint training of person detection and re-identification networks, leading to poor stability in target tracking and efficiency in video surveillance.
Innovation Solution
A person search method based on person re-identification driven localization refinement, incorporating a detection module, region of interest conversion module, and person re-identification module, with supervision using cross-entropy loss and triplet proxy loss, and employing Faster R-CNN and ResNet50 as network backbones to optimize feature extraction and reduce background interference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If joint training by sharing features of person detection and person re-identification networks is used, then the two tasks are performed together, but the accuracy of person search is low due to conflict between shared features
Solution Approach 1:
The patent segments the feature extraction process into two distinct pathways: a detection feature extraction module for person detection and a reidentification feature extraction module for person re-identification. Each module has dedicated feature extractors that process images independently, avoiding the conflict of shared features while maintaining joint training capability. The detection module uses one set of convolutional layers and the reidentification module uses another set, allowing both tasks to be optimized simultaneously without feature interference.
Solution Approach 2:
The patent applies local quality by using different feature extraction configurations for different tasks. The detection feature extractor is optimized for localization with specific convolutional layer settings, while the reidentification feature extractor is optimized for identification with different kernel sizes and pooling operations. This allows each part of the system to have the specific quality needed for its function, resolving the contradiction between joint training and accuracy.
2Measurement precision
If separate detection and re-identification are performed, then the two tasks are performed independently, but joint optimization cannot be carried out well and accuracy is low
Solution Approach 1:
The patent merges the detection and reidentification tasks into a unified network architecture where both modules share the same input processing and loss function optimization. The detection module and reidentification module are trained simultaneously with a combined loss function that optimizes both detection accuracy and reidentification accuracy together. This joint optimization framework allows independent task execution while maintaining coordinated training, resolving the contradiction between independence and joint optimization.
Solution Approach 2:
The patent implements feedback mechanisms where the detection results guide the reidentification process and vice versa. The detection module provides bounding boxes that guide the reidentification module's region of interest extraction, while the reidentification module's confidence scores can feedback to adjust detection priorities. This bidirectional feedback enables joint optimization while maintaining task independence, allowing the system to achieve both goals simultaneously.
3Ease of operation
If detection bounding box with background interference is used, then detection can be performed, but the bounding box is not suitable for person search
Solution Approach 1:
The patent extracts the region of interest from the detection bounding box by removing background interference. After the detection module identifies the person and provides a bounding box, the system extracts only the person region within the bounding box, excluding the surrounding background. This extraction process creates a clean person image suitable for reidentification, resolving the contradiction between maintaining detection capability and improving person search suitability by eliminating background noise.
Data Source
AI summary
The invention discloses a person search method based on person re-identification driven localization refinement. On one hand, the region of interest (ROI) conversion module converts an original input image into a small image corresponding to a ROI, and contradiction existing in part of features shared by a person re-identification network and a detection network is avoided; and on the other hand, loss of the person re-identification network can be transmitted back to the detection network in a gradient manner through the ROI conversion module, the supervision of loss of the person re-identification network for the detection bounding box output by the detection network is realized, and the adjusted detection bounding box can effectively remove background interference, contains more useful attribute information and is more suitable for person search, so that the person search accuracy is greatly improved.