Infrared dim small target tracking method based on multi-scale fusion re-detection mechanism

By employing a multi-scale fusion re-detection mechanism and dynamic template updates, the problems of target loss and background interference in infrared weak target tracking technology are solved, achieving high-precision and stable target tracking results.

CN122199608APending Publication Date: 2026-06-12XIAN UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
XIAN UNIV OF TECH
Filing Date
2026-02-03
Publication Date
2026-06-12

AI Technical Summary

⚠Technical Problem

Existing infrared weak target tracking technology has shortcomings in detection accuracy, response speed and adaptability to complex scenes. In particular, it is difficult to re-lock after the target is lost, background interference affects recognition accuracy, and the lack of a dynamic template update mechanism leads to frequent tracking drift.

⚗Method used

A multi-scale fusion re-detection mechanism is adopted, combined with the SiamFR model, to extract features through Backbone, Neck and head modules, introduce channel and spatial attention, create a trajectory memory bank, and use a re-detection mechanism and dynamic template update strategy to improve tracking robustness.

🎯Benefits of technology

It maintains high robustness in complex scenarios, reduces re-detection efficiency, dynamically adjusts templates to adapt to changes in target appearance, prevents tracking drift, and improves target recognition accuracy and tracking continuity.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122199608A_ABST
    Figure CN122199608A_ABST
Patent Text Reader

Abstract

The application discloses an infrared dim small target tracking method based on a multiscale fusion redetection mechanism, and specifically comprises the following steps: acquiring an infrared image dataset, converting XML annotation information of the dataset into a JSON format, cropping target and search area images from the dataset sequence, screening and grouping the target in each video, and extracting continuous tracking segments; constructing a SiamFR model, inputting a training set into the SiamFR model for training; and inputting a test set into the trained SiamFR model for detection. The method introduces attention and spatial gating, takes into account spatial, channel and semantic features, extracts semantic information and suppresses irrelevant information, and improves tracking performance; a redetection mechanism is used to create a memory bank to save target trajectory information; through a dynamic template updating strategy, tracking drift caused by target appearance changes is avoided, template degradation is prevented, and higher tracking robustness is maintained in a complex scene.
Need to check novelty before this filing date? Find Prior Art