A pedestrian detection algorithm training method based on pedestrian re-identification and text feature matching supervision
By adding pedestrian re-identification and text feature matching branches to the pedestrian detection algorithm model, and combining feature map processing and loss function optimization, the problems of false detection and false negative detection are solved, and the detection accuracy and feature extraction capability are improved.
CN120495978BActive Publication Date: 2026-06-30LINKER
Patent Information
- Authority / Receiving Office
- CN ยท China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- LINKER
- Filing Date
- 2025-04-27
- Publication Date
- 2026-06-30
AI Technical Summary
Technical Problem
Existing pedestrian detection algorithms have a high false detection rate in complex environments and a high false negative rate due to the diversity of pedestrian postures, making it difficult to balance detection accuracy.
Method used
In each stage of the backbone network of the pedestrian detection algorithm model, pedestrian re-identification and text feature matching branches are added. By training and freezing these branches, combined with calibration information and feature map processing, the loss function is calculated to optimize the model parameters.
Benefits of technology
It improves the accuracy of pedestrian detection algorithms, reduces false detection and false negative rates, and enhances the model's sensitivity to and ability to extract features.
โฆ Generated by Eureka AI based on patent content.
Smart Images
Figure CN120495978B_ABST
Abstract
This invention discloses a training method for a pedestrian detection algorithm based on pedestrian re-identification and text feature matching supervision. It adds ReID and text feature matching branches to each stage of the backbone network. During the training process of the pedestrian detection algorithm model, loss calculation is performed using the ReID and text feature matching branches based on manually calibrated coordinates and categories, enhancing the sensitivity of the pedestrian detection model to targets. Furthermore, it compares the predicted target with the manually calibrated information in terms of category and coordinates, performing loss calculation and assigning target weights to each prediction result at each stage, strengthening the extraction of effective target information by the pedestrian detection model and reducing false detections. This scheme is applicable to the training of pedestrian detection algorithms.
Need to check novelty before this filing date? Find Prior Art