.png){kind=icon}
.png){kind=icon}
Return
Target-specific proposals meet memory networks: a unified framework for robust long-term tracking
.png){kind=icon}
PRE
AI
Z
Y
P
DOI：10.1007/s10586-026-06435-9.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Long-term tracking has gained significant attention due to its practical applications. However, existing methods, such as sliding window and re-detection techniques, often generate numerous irrelevant regions unrelated to the target, while frequent target appearance variations further degrade tracking performance. To address these challenges, we propose a unified framework, target-specific proposals meet memory networks, designed for robust long-term tracking, referred to as TSPT. Specifically, TSPT introduces a global search module that leverages swarm intelligence to generate comprehensive region proposals. To refine these proposals, an importance-aware feature fusion strategy is developed to suppress irrelevant candidate regions and enhance target-aware proposal representation. Additionally, we design a complementary module comprising an adaptive short-term memory network (ASMNet) and a verifier. The ASMNet sub-module collects reliable short-term tracking results as template samples to adaptively maintain robust target representations against appearance variations. Meanwhile, the verifier dynamically monitors the reliability of tracking results in real time, enabling seamless switching between local tracking and global search for robust long-term performance. Experimental results on six benchmark datasets demonstrate that the proposed TSPT significantly outperforms several state-of-the-art trackers, and TSPT achieves an area under curve (AUC) score of 62.4% on the LaSOT dataset.
Keywords:
Long-term tracking
Target-specific proposals
Memory networks
Swarm intelligence
Journal
C
IF:
4.1
Papers:
4.8K
Citations:
7.5K
Organization
Cited Papers