Scholarlink

<p>Dragon and lion dances are culturally significant ethnic sports whose preservation and fair competitive judging demand objective, automated action recognition. However, deploying recognition systems in typical performance venues is hindered by limited connectivity, modest hardware, and strict latency constraints. This paper presents Lite-STGCN, a lightweight skeleton-based spatiotemporal graph convolutional network tailored for real-time dragon and lion dance action recognition on edge devices. We first construct a dedicated eight-category dataset with multi-view skeleton annotations. The proposed architecture introduces depthwise separable graph convolutions with an adaptive topology mask that captures task-specific long-range joint couplings, a multi-scale temporal aggregation module with gated fusion to handle the coexisting fast and slow motion regimes characteristic of these performances, and a channel attention compression mechanism to suppress redundant features. A coordinated deployment pipeline integrating structured pruning, quantization-aware training, and TensorRT operator fusion is applied for edge inference. Across five-seed runs, Lite-STGCN achieves 91.7 ± 0.22% top-1 accuracy on the dragon and lion dance dataset with only 1.71&#xa0;M parameters and 0.98 GFLOPs, leading the strongest heavyweight baseline (MS-G3D) and the strongest lightweight baselines (Shift-GCN, EfficientGCN) by statistically significant margins on this in-domain dataset while cutting computational cost by over 40% relative to the ST-GCN baseline; on NTU RGB + D 60 it remains competitive under both cross-subject and cross-view protocols, confirming that the gains are not artefacts of the in-domain split. On the NVIDIA Jetson Xavier NX the fully optimized model reaches 11.3 ms latency at 90.6% accuracy within a 7.1&#xa0;W power envelope, which meets competition-side real-time needs; on the more limited Jetson Nano it sustains only near-real-time throughput, better suited to coaching review than to live scoring.</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

Lite-STGCN: a lightweight spatiotemporal graph convolutional network for real-time dragon and lion dance action recognition on edge devices

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月4日

查看全文
  • Maobin Zhu

摘要

Dragon and lion dances are culturally significant ethnic sports whose preservation and fair competitive judging demand objective, automated action recognition. However, deploying recognition systems in typical performance venues is hindered by limited connectivity, modest hardware, and strict latency constraints. This paper presents Lite-STGCN, a lightweight skeleton-based spatiotemporal graph convolutional network tailored for real-time dragon and lion dance action recognition on edge devices. We first construct a dedicated eight-category dataset with multi-view skeleton annotations. The proposed architecture introduces depthwise separable graph convolutions with an adaptive topology mask that captures task-specific long-range joint couplings, a multi-scale temporal aggregation module with gated fusion to handle the coexisting fast and slow motion regimes characteristic of these performances, and a channel attention compression mechanism to suppress redundant features. A coordinated deployment pipeline integrating structured pruning, quantization-aware training, and TensorRT operator fusion is applied for edge inference. Across five-seed runs, Lite-STGCN achieves 91.7 ± 0.22% top-1 accuracy on the dragon and lion dance dataset with only 1.71 M parameters and 0.98 GFLOPs, leading the strongest heavyweight baseline (MS-G3D) and the strongest lightweight baselines (Shift-GCN, EfficientGCN) by statistically significant margins on this in-domain dataset while cutting computational cost by over 40% relative to the ST-GCN baseline; on NTU RGB + D 60 it remains competitive under both cross-subject and cross-view protocols, confirming that the gains are not artefacts of the in-domain split. On the NVIDIA Jetson Xavier NX the fully optimized model reaches 11.3 ms latency at 90.6% accuracy within a 7.1 W power envelope, which meets competition-side real-time needs; on the more limited Jetson Nano it sustains only near-real-time throughput, better suited to coaching review than to live scoring.