Single Target Tracking With Contextual Prompts in Scene Understanding
Xuelin Liu, Huiyu Zhou, Junyu Dong, Jingjing Xiao, Xinghui Dong · IEEE Transactions on Artificial Intelligence · 2025
This paper presents a single object tracker that incorporates scene understanding, referred to as SU-STTrack, which explicitly encodes the contextual information by the given or Multi-modal Large Language Model (MLLM) generated descriptions of the tracking scene. SU-STTrack merges the linguistic contextual prompts encoded using a pre-trained LLM and visual features using a dual attention mechanism to strengthen tracking robustness and adaptability, where channel attention blocks are fixed within a multi-modality feature fusion process. The experience replay strategy is further proposed, to maintain long-term tracking performance by periodically refreshing the tracking template with accumulated experiences, preventing the model from catastrophic forgetting. SU-STTrack enables both vision-only and vision-language tracking tasks to share the same parameters. Through extensive experiments on vision-language data sets (TNL2k, LaSOT and LaSOText) and vision-only data sets (UAV123, NfS and OTB100), our tracker achieves 0.569 on TNL2K, 0.628 on the challenging LaSOT, 0.528 on LaSOText, 0.646 on UAV123, 0.603 on NfS and 0.718 on OTB100 in terms of the Area Under the Curve (AUC) metric with the inference speed of 36.3 FPS. SU-STTrack normally surpasses state-of-theart methods and shows superior generalization ability across diverse and challenging tracking scenarios11The code and models are available athttps://indtlab.github.io/projects/SUSTTrack..