Memory optimization network for enhanced vision-language target tracking

Jianwei Zhang

doi:10.1007/s11227-026-08377-w

Return

Memory optimization network for enhanced vision-language target tracking

2026-03-04

PRE

DOI：10.1007/s11227-026-08377-w

Original

Original request for help

Save

Abstract

En 中文

In video-language tracking, providing comprehensive and precise historical information is crucial to handle scenarios involving target appearance variations. However, the historical information utilized by existing methods contains substantial noise and typically requires repetitive training. It leads to imprecise historical reference and introduces considerable computation, thus entailing degraded accuracy as well as increased complexity of the model. To address this issue, we propose a Memory Optimization Network for Enhanced Vision-Language Target Tracking (VLMOTrack). Specifically, we design a two-stage token filtering mechanism (TTFM). It employs attention hierarchical for adjusting target weights to achieve progressive screening. The first stage uses the stable initial template features for coarse-grained filtering background information. The second stage finely eliminates non-target tokens by utilizing steady language and rich historical visual prompts. Thus, the model obtains precise target feature representations. Furthermore, we construct an adaptive memory prompt generation mechanism (AMPGM). It employs a convolution-based selection strategy to identify reliable frames, which are encoded into a key-value pair structure through a lightweight encoder and convolutional block for storage in a memory bank. Subsequent frames adaptively aggregate values, thereby generating accurate and customized memory prompts. Finally, we propose a memory prompt propagation strategy. The memory prompts are fused with search features through a convolution block to enhance the temporal consistency and discriminability. It then operates as described in the TTFM, thereby propagating historical prompts to subsequent frames for prediction and inference. We conduct experiments on the TNL2K, LaSOT, OTB99-Lang and LaSOText datasets. The experimental results demonstrate that VLMOTrack possesses favorable competitiveness.

Keywords:

Object tracking

Vision-language

Memory network

Prompt learning