Scholarlink

<p>High-speed tennis tracking is constrained by severe motion blur and the strict real-time requirements of edge-side deployment. To address these challenges, this paper proposes TGMA-Net, a lightweight, high-precision framework built upon a coarse-to-fine tracking paradigm. First, we construct an efficient encoder using an EGhostNet backbone and a double motion frame differencing (Double-FD) module, which explicitly encodes motion polarity to provide the network with real-time direction-aware priors. Second, to suppress noise propagation and background clutter, we integrate attention gates (AG) within the skip connections, which utilize deep semantic signals to perform spatial filtering on shallow-level features. Finally, we introduce an LKA-based trajectory refiner, which employs a large receptive field to rectify localization drift and shape dispersion caused by high-speed movement. Combined with a multi-scale deep supervision strategy, extensive experiments demonstrate that TGMA-Net requires only 1.87&#xa0;M parameters (an 81.6% reduction compared to TrackNetV2) while achieving a real-time inference speed of 173&#xa0;FPS. Validated across multi-scenario datasets, our model significantly outperforms heavy-weight baselines with a Recall of 97.80% and achieves a sub-pixel RMSE of 1.0515 pixels, offering a robust and efficient solution for real-time sports tactical analysis in complex environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

TGMA-Net: a lightweight spatiotemporal attention network for precise real-time tennis tracking

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Lin Meng,
  • Wenping Li,
  • Jijun Tong

摘要

High-speed tennis tracking is constrained by severe motion blur and the strict real-time requirements of edge-side deployment. To address these challenges, this paper proposes TGMA-Net, a lightweight, high-precision framework built upon a coarse-to-fine tracking paradigm. First, we construct an efficient encoder using an EGhostNet backbone and a double motion frame differencing (Double-FD) module, which explicitly encodes motion polarity to provide the network with real-time direction-aware priors. Second, to suppress noise propagation and background clutter, we integrate attention gates (AG) within the skip connections, which utilize deep semantic signals to perform spatial filtering on shallow-level features. Finally, we introduce an LKA-based trajectory refiner, which employs a large receptive field to rectify localization drift and shape dispersion caused by high-speed movement. Combined with a multi-scale deep supervision strategy, extensive experiments demonstrate that TGMA-Net requires only 1.87 M parameters (an 81.6% reduction compared to TrackNetV2) while achieving a real-time inference speed of 173 FPS. Validated across multi-scenario datasets, our model significantly outperforms heavy-weight baselines with a Recall of 97.80% and achieves a sub-pixel RMSE of 1.0515 pixels, offering a robust and efficient solution for real-time sports tactical analysis in complex environments.