一种目标跟踪方法、终端设备及存储介质
By introducing a spatial attention mask matrix and a channel attention mechanism, and combining a multimodal target template pool to optimize the correlation filter, the tracking failure problem of the discriminative correlation filtering method in occluded scenarios is solved, achieving higher tracking accuracy and robustness.
CN115272707BActive Publication Date: 2026-07-17MINNAN NORMAL UNIV
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- MINNAN NORMAL UNIV
- Filing Date
- 2022-07-28
- Publication Date
- 2026-07-17
AI Technical Summary
Technical Problem
The discriminant correlation filtering method failed to effectively evaluate the reliability of target feature channels and spatial features in occluded scenarios, leading to the failure of the tracking algorithm in occluded scenarios.
Method used
By introducing a spatial attention mask matrix and a channel attention mechanism, combined with a multimodal target template pool, the learning and updating of relevant filters are optimized to achieve robustness in target tracking.
Benefits of technology
It improves the accuracy and robustness of target tracking in occluded scenarios, reduces the possibility of target loss, and enhances the tracking algorithm's resistance to occlusion.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN115272707B_ABST
Abstract
本发明涉及一种目标跟踪方法、终端设备及存储介质，该方法中包括：当接收到新一帧的图片时，提取特征并估计空间注意力掩码矩阵；获取当前帧样本各通道响应，并更新相关滤波器中各通道的实际注意力权重，计算当前帧各样本的相关滤波响应，并提取其中响应最大的样本作为当前帧的候选样本；提取候选样本的特征，并将其与历史多模态目标模板池内各目标模板的特征进行相似度计算，当计算的最大相似度超过相似度阈值时，判定该候选样本可靠，将候选样本对应的图片添加至历史多模态目标模板池内，更新当前帧训练样本、滤波器、模板池、目标和背景颜色直方图；否则，不更新。本发明可以实现遮挡场景下鲁棒跟踪的目的，提升了目标跟踪的准确性。
Need to check novelty before this filing date? Find Prior Art