Efficient Plug-and-Play Mamba-based Selective Target State Modeling for Lightweight Visual Tracking

Temporal target state modeling is essential for robust visual tracking, yet existing approaches often rely on deep or iterative architectures that incur excessive computational and parametric costs, making them unsuitable for lightweight real-time deployment. To address this challenge, we propose TSTrack, the first lightweight tracking framework that integrates a State Space Model (SSM). TSTrack consists of two innovative modules: (1) Target-Aware Mamba (TAM): It enables low-latency adaptive temporal modeling by fusing real-time search states, static templates, and compressed historical context in Mamba's hidden states through selective bidirectional interaction. As a plug-and-play module, TAM improves the performance of existing lightweight trackers with minor computational and parameter increase (approximately +6% parameters and +1% MACs); (2) Spatial-Channel Aggregation Module (SCAM): It hierarchically refines target features via dual-path attention that coordinates spatial activation enhancement and channel-wise feature recalibration, boosting localization precision in complex scenarios. Our TSTrack achieves promising tracking speed with competitive tracking performance. For instance, it suppresses previous lightweight tracking methods across six commonly used tracking benchmarks, such as LaSOT, GOT-10k, and TrackingNet. Simultaneously, it can run at 41_fps on CPU devices. This work redefines efficiency-accuracy trade-offs in lightweight visual tracking and advances temporal modeling theory in resource-constrained computer vision tasks.