Online dense point tracking system and method based on feature projection

By introducing feature projection and memory bank update mechanisms in the video processing system, the problems of low long-distance motion capture efficiency and appearance drift in real-time streaming videos in the prior art are solved, and high-precision and high-efficiency online dense point tracking are achieved.

CN120198460APending Publication Date: 2025-06-24FUDAN UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510196030.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-02-21
Publication Date
2025-06-24

AI Technical Summary

Technical Problem

The prior art is difficult to fully capture dense long-distance motion in real-time streaming video, and there are problems of appearance drift and low occlusion processing efficiency.

Method used

A online dense point tracking system based on feature projection is designed, including an image encoder, feature enhancement unit, short-term perception memory unit, point tracking decoder, feature projection unit and memory bank. Through feature projection and memory bank update, dense features are enhanced and long-distance point tracking are achieved.

Benefits of technology

It alleviates the appearance drift problem of point tracking, improves the accuracy and efficiency of online real-time point tracking, and can effectively deal with long-distance and short-term occlusion.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120198460A_ABST
    Figure CN120198460A_ABST
Patent Text Reader

Abstract

The invention relates to an online dense point tracking system and method based on feature projection. The system comprises an image encoder used for extracting dense features from a video frame; the feature enhancement unit is used for reading the key value pairs from the memory bank so as to perform enhancement processing on the dense features to obtain enhanced features; the short-term perception memory unit is used for perceiving motion memory features; the point tracking decoder is used for decoding to obtain a point tracking corresponding relation and a visibility mask; the feature projection unit performs forward projection on the features of the first frame in the video by using the point tracking corresponding relation and the visibility mask, and transmits a projection result to the memory bank for updating; the memory bank comprises a key queue and a value queue and is used for updating and storing the query vector and the projection result. Compared with the prior art, the method has the advantages that a memory mechanism is introduced, appearance information is spread through feature projection, short-term motion dynamics are modeled, the appearance drift problem of point tracking can be relieved, and the precision and efficiency of online real-time point tracking are improved.
Need to check novelty before this filing date? Find Prior Art