Fall-Detection - Proof of Concept

What’s done

• Built a working end-to-end app on the Voyager SDK: camera > YOLOv8-Pose > decode > filters > overlay > action.

• Added OpenCV overlay (boxes + 17-key point skeleton) and optional MP4 recording.

• Integrated a Kasa plug trigger, I will be configuring an Alexa routine once I figure out some issues.

• Here is a screen recording of the fall detection working:  

   

Pipeline YAML (high level)

• Model: yolov8lpose-coco 

• Preprocess: letterbox to 640×640, torch-totensor.

• Postprocess: decodeyolopose with NMS + confidence thresholds; outputs scaled back to original frame size.

• Outputs land in frame.metao'yolov8lpose-coco'] as:

  • boxes (N×4, xywh), keypoints (N×17×3), scores (N, pose conf).

How I decide what a “fall” is (current rules)

• Quality gates (to ignore furniture/noise):

  • Min box area ≈ 90,000 px², max ≈ 250,000 px².

  • Min pose confidence 0.65.

  • At least 8 visible keypoints with per-keypoint confidence ≥ 0.35.

• Fall heuristic: box goes wide>tall (high w/h) and vertical spread of keypoints is low (lying posture).

• Temporal smoothing: require 5 consecutive fallen frames before triggering;  

• Action: pulse Kasa plug (default 5s) 

NEW / In progress: “lying on couch/bed” detection

• Goal: distinguish fallen on floor vs lying safely on a surface (couch/bed) to reduce false alarms.

• Approach (rule-based first, then refine):

  • No descent: low vertical velocity before the posture change (no sudden drop).

  • Support height: head/hips above floor by a threshold (e.g., >15% of frame height), consistent with a raised surface.

  • Horizontal posture, but with stable box bottom near a calibrated couch/bed ROI (simple per-site calibration).

  • Longer dwell for “lying” state without triggering Kasa; label as “resting” instead.

I’m really excited with where I’m at with this project. My initial goals are met and I’m actually able to refine the logic before the deadline to get a much better prototype working!