.png){kind=icon}
.png){kind=icon}
Return
Target prompt token-based lightweight transformer for single object tracking algorithm
.png){kind=icon}
PRE
AI
DOI：10.37188/cjlcd.2026-0058.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Although Transformer-based trackers have demonstrated strong accuracy owing to their superior global modeling capability, they still suffer from large model size and high computational latency. To address these issues, this paper proposes TPTTrack, a lightweight Transformer-based single object tracking method built upon target prompt tokens. Specifically, a target prompt token is introduced and jointly fed with the target template and search-region features into a lightweight Transformer backbone, where target priors are exploited to guide cross-region feature interaction and enhance tracking robustness. In addition, a hierarchical attention decoupling mechanism is designed to perform feature extraction and fusion in shallow layers while reducing redundant self-attention computation in deep layers, thereby lowering the overall computational cost. A lightweight autoregressive sequence prediction module is further incorporated to enable dynamic target-state modeling and efficient state estimation. Comparative experiments on the GOT-10k and LaSOT benchmarks show that TPTTrack improves Average Overlap on GOT-10k by 0. 9% and precision on LaSOT by 1. 4% over the strongest real-time baseline. Meanwhile, the model contains only 5. 5M parameters and 1. 79G FLOPs, and achieves inference speeds of 111. 6 FPS on GPU and 20.76 FPS on CPU, with the CPU speed being 10. 8% higher than that of the strongest real-time baseline. Overall, compared with existing lightweight tracking algorithms, the proposed method achieves superior performance in terms of accuracy, speed, and deployment friendliness, providing an effective lightweight solution for single object tracking in resource-constrained scenarios.
Keywords:
single object tracking
transformer
lightweight model
target prompt token
hierarchical attention decoupling
Journal
C
IF:
0.7
Papers:
58
Citations:
429