> For the complete documentation index, see [llms.txt](https://docs.augelab.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.augelab.com/function-blocks/blocks-reference/ai-blocks/hand-pose-estimation.md). # Hand Pose Estimation This function block detects and estimates hand keypoints (21 per hand) in images and provides both a visual overlay and structured detection data. It is designed for real-time use and offers controls for detection sensitivity, keypoint visibility, style of skeleton output, and the maximum number of hands to process. ## 📥 Inputs `Image` Feed an image (camera frame, loaded image, or processed image) to analyze for hands. ## 📤 Outputs `Visualization` Annotated image showing keypoints, skeletons and bounding boxes. `Hands` Structured detection data (list/dictionary) including bounding boxes, per-keypoint positions, confidence scores and visibility flags. `Model Info` Basic run-time information such as selected skeleton style and threshold settings. `Hand Count` Number of hands detected (after applying the configured limits). ## 🕹️ Controls `Skeleton Style` Choose how keypoints/skeletons are formatted for the visualization (e.g., MMPose or OpenPose style). `Det Threshold` Adjust minimum confidence required for a hand detection to be considered valid (0–100 scale). `Keypoint Threshold` Set the minimum confidence for an individual keypoint to be considered visible (0–100 scale). `Max Hands` Limit how many hands are returned and drawn (useful to reduce output size and processing for crowded scenes). ## 🎨 Features * Visual overlay with keypoints, skeleton connections and bounding boxes for each detected hand. * Structured JSON-like output for downstream logic: bounding boxes, per-keypoint (x,y) positions, confidence and visibility. * User-adjustable thresholds to trade off sensitivity vs. false positives. * Limit the number of hands processed with `Max Hands` for predictable downstream behavior. * Automatically uses available hardware to improve speed (will prefer GPU if available). ## 📝 Usage Instructions 1. Provide an image source into the `Image` input (live camera, stream, or image file). 2. Choose preferred `Skeleton Style` for visualization and downstream format. 3. Adjust `Det Threshold` to control whether weak detections are ignored. 4. Adjust `Keypoint Threshold` to control which keypoints are considered visible. 5. Set `Max Hands` if you only want to track a limited number of hands. 6. Read outputs: use `Visualization` to preview, and use `Hands` / `Hand Count` for logic, UI or logging. ## 📊 How it runs When provided with an image, the block analyzes the picture for hands, applies detection and keypoint confidence thresholds, limits results by `Max Hands`, and then outputs: a visual image with overlays, a structured list of detected hands with bounding boxes and per-keypoint details, a small model info summary, and the number of detected hands. ## 💡 Tips and Tricks * For live input combine with `Camera USB`, `Camera IP (ONVIF)`, or `Stream Reader` to feed continuous frames. * Use `Show Image` to preview the `Visualization` output in a larger window while tuning thresholds. * Preprocess noisy images with `Blur`, `Denoising` or `Image Resize` to improve detection stability. * If the hands appear cropped or you only want to analyze a specific area, place an `Image ROI Select` or `Image ROI` block before this block. * To annotate results for reporting, combine `Visualization` with `Write Text On Image` or `Draw Result On Image` and then save with `Image Logger`, `Image Write` or `Record Video`. * Use `Object Detection` or `Object Detection - Custom` before this block when you want to first locate people and then analyze only person regions for hands—this reduces false positives and speeds processing. * If you need full-body keypoints as well as hand keypoints, consider pairing with `Skeleton Estimation` and merge results in subsequent processing steps. ## 🛠️ Troubleshooting * No detections: Try lowering `Det Threshold` and `Keypoint Threshold` slightly, or improve image clarity with `Image Resizer`. * False positives / noisy keypoints: Increase thresholds and/or crop the region of interest with `Image ROI Select` to remove clutter. * Too slow: Lower image resolution via `Image Resize`, reduce `Max Hands`, or use a faster image source. Using a system with GPU will accelerate processing. * Missing dependencies or model not available: The block requires the hand-pose model to be available. If the model or runtime components are not present, follow the application’s module installer / module downloader to add the required runtime and model packages. If you need example combinations or a recommended small pipeline for live hand tracking (camera → preprocess → hand pose → display / save), ask for a suggested block chain and a short explanation.