AMPTrack: Tracking Objects using Appearance and Motion Prompts

doi:10.1109/tcsvt.2026.3718656

AMPTrack: Tracking Objects using Appearance and Motion Prompts

2026-07-30

0

PRE

AI

DOI：10.1109/tcsvt.2026.3718656

Original

Original request for help

Save

Abstract

En 中文

Appearance adaptation and motion adaptation are critical components in point cloud tracking. However, the lack of informative prompts in existing methods limits their adaptability to varying scenarios, thereby resulting in suboptimal tracking performance. To overcome these limitations, we propose AMPTrack, which enhances target tracking by leveraging appearance and motion prompts. Specifically, we introduce a prompt-enhanced adjacent transformer (PAT) module, which propagates appearance, motion, and mask cues from adjacent frames to the current frame using a triple-decoupling and double-injection approach. And it injects appearance and motion prompts into the current features through cross-attention, thereby providing explicit guidance for modeling. Additionally, to obtain more comprehensive appearance and motion prompts, we propose an incessant appearance prompt (IAP) module and an incessant motion prompt (IMP) module. These modules aggregate historical appearance and trajectory cues into appearance and motion prompts, respectively, using a selective scanning mechanism, and update both prompts online. Extensive experiments on the KITTI, NuScenes, and Waymo datasets demonstrate that our method achieves state-of-the-art (SOTA) performance.

Keywords:

Point cloud object tracking

appearance prompt learning

motion prompt learning

AMPTrack: Tracking Objects using Appearance and Motion Prompts

Abstract

Journal

Organization

Cited Papers