Sample generation method and device for trajectory prediction, computer equipment and medium

By acquiring and processing the trajectories of FPV and BEV video streams, samples carrying perceptual defect information are generated, which solves the problem of decreased prediction accuracy caused by domain offset in machine learning models and improves the robustness and accuracy of trajectory prediction.

CN122090197APending Publication Date: 2026-05-26HONG KONG UNIV OF SCI & TECH (GUANGZHOU)
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
HONG KONG UNIV OF SCI & TECH (GUANGZHOU)
Filing Date
2026-01-05
Publication Date
2026-05-26

AI Technical Summary

Technical Problem

In existing trajectory prediction research, machine learning models exhibit a significant domain shift problem between the training data environment and the real deployment environment, leading to a decrease in prediction accuracy.

Method used

By acquiring first-person perspective FPV video streams and bird's-eye view BEV video streams from the same acquisition device, multiple trajectories are extracted, and target trajectory pairs are obtained by pairing based on multi-order motion matching costs. Samples carrying perception defect information are generated for training machine learning models.

Benefits of technology

It effectively mitigates the domain offset problem and improves the trajectory prediction accuracy of machine learning models, especially significantly enhancing matching robustness in crowded or occluded scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122090197A_ABST
    Figure CN122090197A_ABST
Patent Text Reader

Abstract

The invention discloses a sample generation method and device for track prediction, computer equipment and a medium, and belongs to the technical field of artificial intelligence. The method comprises the following steps: acquiring a first-person view angle FPV video stream and a bird's-eye view angle BEV video stream which are synchronously acquired by at least two image sensors arranged on the same acquisition equipment; wherein the FPV video stream carries perception defect information; extracting a plurality of first tracks from the FPV video stream, and extracting a plurality of second tracks from the BEV video stream; determining a plurality of candidate trajectory pairs according to the first trajectories and the second trajectories; determining the multi-order motion matching cost of each candidate track pair; based on each multi-order motion matching cost, performing pairing to obtain a target trajectory pair; wherein the target track pair is used for generating a sample, and the sample comprises the perception defect information. The prediction accuracy of the machine learning model can be improved.
Need to check novelty before this filing date? Find Prior Art