Mamba meets tracker: exploiting token aggregation and diffusion for robust unmanned aerial vehicles tracking

Abstract

The Transformer-based tracking approach achieves excellent results in unmanned aerial vehicles (UAV) tracking tasks. However, the existing tracking framework usually deals with this problem by visual grounding and visual tracking separately. This independent framework does not consider the correlation between the two steps mentioned above, that is, natural language description can provide global semantic information. Meanwhile, a separate framework is unable to conduct end-to-end training. As a remedy, We propose a joint natural language Mamba based tracking framework (named TADMT). Specifically, we propose a token aggregator that condenses rich features into a small number of visual tokens through a coarse to fine strategy to improve subsequent tracking speed. Then, we designed a mamba module based on the serpentine scanning strategy to effectively establish the relationship between natural language and visual images. In addition, we have designed a novel shift add multilayer perceptron in the prediction head, with the aim of achieving final classification and localization with less computation. Numerous experimental results have shown that TADMT achieves good tracking performance on six UAV tracking datasets and three general tracking datasets, with an average speed of 120FPS. The experimental results on the embedded platform also demonstrate the applicability of TADMT on UAV platforms.