搜索

PythonReactSEOPDF
GitHub Skill

tune-detection

Tune exercise detection sensitivity and thresholds. Use when exercises aren't counting correctly, have false positives, or need calibration adjustments.

Flow-Club☆ 3 仓库 Stars0 仓库 Forks

安装与使用

请根据来源仓库和你的 AI 工具说明安装此技能。先阅读技能文件和依赖要求,再添加到工作流。

查看来源 ↗使用指南

技能说明

Tune Detection Thresholds

Common Issues

Problem Likely Cause Fix
Not counting reps Thresholds too strict Lower down threshold or raise up threshold
Double counting Thresholds too loose Tighten thresholds, add hysteresis
Counts on wrong motion Wrong landmarks Check landmark IDs match exercise
Works for some people Fixed thresholds Use body-relative thresholds with getBodyScale()

Threshold Locations

JSON configs (preferred):

exercises/*.json → detection.thresholds

Legacy functions in exercise_ui.html:

  • detectLegacySquat - DOWN_ANGLE, UP_ANGLE
  • detectLegacyPushup - DOWN_ANGLE, UP_ANGLE
  • detectLegacyJumpingJack - arm position relative to shoulders
  • detectLegacyCalfRaise - legLength * 0.04 threshold
  • detectLegacySideStretch - bodyScale * 0.4 threshold

Detection Helpers

The codebase includes helpers for robust detection:

// Check if landmarks are visible
isVisible(landmarks, [23, 24, 25, 26], 0.5)

// Normalize to body size
const bodyScale = getBodyScale(landmarks);
const threshold = baseThreshold * bodyScale;

Testing Changes

  1. Edit threshold in JSON or HTML
  2. Restart tracker: ./exercise_tracker.py user_prompt_submit '{}'
  3. Watch the status text - it shows live angle/distance values
  4. Adjust based on when state transitions happen

Both-sides Averaging

For angle-based exercises, use joint_alt in JSON to average both sides:

"landmarks": {
  "joint": [23, 25, 27],
  "joint_alt": [24, 26, 28]
}

使用示例

查看来源仓库中的使用示例与配置说明。

查看来源 ↗

来源历史

收录时间
2026-01-20
索引更新时间
2026-02-02
Branch
main

收录与同步日期属于本站记录。源码修改历史请查看来源仓库。

GitHub ↗

chat Comments (0)

chat_bubble_outline

No comments yet. Be the first to share your thoughts!