← All projects

ArcheryEdge

A stereo computer vision system that analyzes archery form in real time and delivers coaching feedback on-device.

System architecture

Two Arducam UC-619 CSI cameras feed synchronized frames into a Jetson Orin Nano Super. YOLO26-pose runs with TensorRT FP16 for 2D keypoint detection in each view. Stereo calibration uses a ChArUco board (6×5 squares, DICT_4X4_50), and 3D joint positions are computed via DLT triangulation with OneEuroFilter smoothing.

Coaching pipeline

Posture metrics extracted from the 3D skeleton feed into a RAG system built on FAISS and sentence-transformers, which retrieves relevant coaching knowledge. A Qwen2.5-3B model running on-device generates natural-language feedback. A separate Raspberry Pi Zero 2W camera on the target face provides shot grouping context.

Key challenges solved

  • Jetson dependency resolution (PyTorch wheel, NumPy, torchvision compatibility)
  • Stereo calibration accuracy at archery-relevant distances (15–70m)
  • Real-time inference budget on edge hardware
  • Smooth 3D tracking without jitter for form analysis