An intelligent analysis method for transient events based on multi-modal spatio-temporal feature fusion

By using a multimodal spatiotemporal feature fusion method, combined with depth model and gray-scale difference analysis, the problems of missed detection and inaccurate positioning of sudden events in UAV monitoring videos were solved, achieving high-precision and stable instantaneous event detection and positioning.

CN122115502APending Publication Date: 2026-05-29CHANGGUANG SATELLITE TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
CHANGGUANG SATELLITE TECH CO LTD
Filing Date
2026-02-10
Publication Date
2026-05-29

AI Technical Summary

Technical Problem

Existing drone surveillance videos struggle to capture heterogeneous features simultaneously in detecting sudden events, leading to missed detections and false detections. Furthermore, they lack accurate estimation of event locations, making it difficult to distinguish between real events and background interference, especially in complex environments.

Method used

A multimodal spatiotemporal feature fusion method is adopted, which combines deep model detection and inter-frame grayscale difference analysis. The video is stabilized by optical flow tracking and smooth incremental processing. Semantic features are extracted using the YOLOv5 model and combined with adaptive threshold and Gaussian consistency prior term for event localization.

Benefits of technology

It achieves reliable detection and localization of transient events in complex environments, with a prediction recall rate of 100%, a relative time error reduced to 75.876 milliseconds, and a relative position error controlled at 8.2864%, significantly improving detection accuracy and stability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122115502A_ABST
    Figure CN122115502A_ABST
Patent Text Reader

Abstract

The application relates to a kind of instantaneous event intelligent analysis methods based on multi-modal spatiotemporal feature fusion.The application relates to a kind of instantaneous event intelligent analysis methods based on multi-modal spatiotemporal feature fusion.The application first uses a light-weight model to quickly identify possible event candidate regions, and simultaneously enhances the sensitivity to weak events by combining the brightness jump signal; subsequently, high-precision optical flow estimation and local texture change analysis are used to separate event-related local anomalies from dynamic backgrounds; in the time sequence dimension, the occurrence time of the event is accurately determined by fusing multiple clues such as firelight appearance, brightness peak and optical flow enhancement; further, the stable positioning of the event center is realized by combining the event region centroid trajectory, time sequence brightness distribution and geometric constraints. The overall process does not require large-scale training data, and still maintains good robustness in complex environments, and is suitable for multi-type sudden event analysis in monitoring video and unmanned aerial vehicle inspection.
Need to check novelty before this filing date? Find Prior Art