LLLMWIKI
ArticleResearch mapReading portalMetadata

Source note · Updated 2026-07-12

MonoTrackShuttle Trajectory Reconstruction from Monocular Badminton Video

A modular monocular pipeline combining court, pose, shuttle tracking, hit segmentation, and physics-constrained 3D reconstruction.

#near-cvpr-2025#video-understanding#tracking#primary-source

一句话结论

MonoTrack 将 court、pose、2D shuttle tracking、HitNet 与带空气阻力的非线性优化串成固定广播视角单打的单目 3D 轨迹系统；它直接覆盖羽毛球四层中的前三层，但真实比赛没有 3D GT，端到端误差和适用范围必须谨慎解释。

数据与 split

  • TrackNetV2 数据：26 场国际单打、约 77K frames；3 场 test matches 约 12K frames。
  • 其余数据按 frame 90/10 train/val，可能把同 rally 相邻帧跨集合，验证集有近重复泄漏风险；test match 独立。
  • 另标 40 场 YouTube 视频测试 court detection；另造 10K synthetic 3D trajectories。
  • 只验证 singles、fixed overhead/static broadcast-view。

方法

  1. Hough lines + graph-based bipartite partition 检测 court；
  2. HRNet pose 与 3-in/3-out modified U-Net 检测 shuttle；
  3. 约 16K 参数双层 GRU HitNet 融合 court/pose/shuttle，并以击球间隔和双方交替规则做动态规划；
  4. 每拍使用 drag dynamics、camera reprojection、击/接球方先验与出界惩罚优化 3D trajectory。

核心结果与消融

  • Court success 73.9%→85.5%，IoU 0.96→0.97，速度快 40×。
  • Shuttle detector 2.9M vs 11.3M parameters，accuracy 88.6% vs 84.0%。
  • HitNet：Shuttle only F1 0.815；+Pose 0.850；+Court 0.866；再加 constrained optimization 0.946（recall 94.3%、accuracy 89.7%）。
  • Synthetic 3D error：reprojection-only 14.9 cm，full priors 8.0 cm；synthetic 由同一物理模型生成和拟合，偏乐观。
  • Real 只报告 2D reprojection：GT shuttle/hits 8.1/8.9/9.8 px，全自动 37.1/28.8/23.3 px。
  • 约 90% hit accuracy 仍可导致约 20% shots 重建错误，因为一个 hit error 会破坏相邻两拍。

成本与局限

HitNet 可达每秒数千帧，但论文没有 end-to-end FPS、硬件或 HRNet/非线性优化 latency；不能据此声称全系统实时。Real data 无 3D GT；评测排除常遮挡 serve 与未标 landing；court failure、出画、hit 错误会级联。Synthetic speed “1500 m/s” 与现实和模型约束冲突，复现前需核对。

对羽毛球四层 pipeline 的价值

  • Shuttle tracking：直接证据，但 88.6% frame accuracy 不等于完整 trajectory。
  • Court/3D reconstruction：直接 baseline；真实 3D accuracy 未证明。
  • Hit/phase alignment：直接证据，HitNet F1 0.946；仅单打固定视角。
  • Action quality assessment：无 stroke quality、技术错误或教练反馈证据。

Demo 应保留 observed/interpolated 点标记、hit/court confidence 与 3D residual，并补 match-level split、serve/landing coverage、真实多视角 3D GT。

深分析与原始材料

  • raw/ingest/2026-05-05-monotrack-shuttle-trajectory-reconstruction/analysis.md
  • raw/ingest/2026-05-05-monotrack-shuttle-trajectory-reconstruction/paper.pdf
  • raw/ingest/2026-05-05-monotrack-shuttle-trajectory-reconstruction/paper-text.md
  • raw/ingest/2026-05-05-monotrack-shuttle-trajectory-reconstruction/links.yaml

相关页面

Metadata
{
  "id": "2026-05-05-monotrack-shuttle-trajectory-reconstruction",
  "type": "source",
  "title": "MonoTrack：单目羽毛球击球分割与 3D 轨迹重建",
  "title_en": "MonoTrack: Shuttle Trajectory Reconstruction from Monocular Badminton Video",
  "summary_en": "A modular monocular pipeline combining court, pose, shuttle tracking, hit segmentation, and physics-constrained 3D reconstruction.",
  "status": "reviewed",
  "created": "2026-05-05",
  "updated": "2026-07-12",
  "tags": [
    "near-cvpr-2025",
    "video-understanding",
    "tracking",
    "primary-source"
  ],
  "note_status": "reviewed",
  "source_type": "paper",
  "venue": "CVPR Workshop 2022",
  "authors": [
    "Paul Liu",
    "Jui-Hsien Wang"
  ],
  "published_at": "2022-04-04",
  "ingested_at": "2026-05-05",
  "canonical_links": [
    "https://openaccess.thecvf.com/content/CVPR2022W/CVSports/html/Liu_MonoTrack_Shuttle_Trajectory_Reconstruction_From_Monocular_Badminton_Video_CVPRW_2022_paper.html",
    "https://openaccess.thecvf.com/content/CVPR2022W/CVSports/papers/Liu_MonoTrack_Shuttle_Trajectory_Reconstruction_From_Monocular_Badminton_Video_CVPRW_2022_paper.pdf",
    "https://arxiv.org/abs/2204.01899"
  ],
  "raw_entry": "raw/ingest/2026-05-05-monotrack-shuttle-trajectory-reconstruction/",
  "topics": [
    "topics/sports-ai-video-understanding",
    "topics/sports-ai-roadmap",
    "topics/video-understanding"
  ],
  "entities": [
    "entities/sportsmot"
  ],
  "claims": [],
  "questions": []
}