ray96nex/yolo-athletic-pose-estimation

Custom YOLO Pose Estimation for Sprint Analysis

Overview

This repository contains a custom-trained YOLO pose estimation model designed specifically for athletic movement analysis, with a focus on sprint biomechanics and stride detection.

The model extends the COCO 17-keypoint schema and applies temporal smoothing during inference for slow-motion footage. It is optimized for single-person side-view videos.

---

Model Details

• Framework: Ultralytics YOLOv11 Pose
• Model variant: yolo11n-pose
• Input size: 640×640
• Training epochs: 100
• Device: CPU
• Precision: FP32
• Pretrained: Yes

---

Key Features

• Side-view sprint video optimized
• Slow-motion analysis
• Single-person assumption
• Temporal smoothing compatible
• Exportable to ONNX and TorchScript

---

Performance Metrics

Metrics computed on side-view slow-motion sprint clips:

{
  "detection_metrics": {
    "precision": 0.995,
    "recall": 0.952,
    "mAP50": 0.979,
    "mAP50-95": 0.938
  },
  "pose_metrics": {
    "precision": 0.500,
    "recall": 0.488,
    "mAP50": 0.493,
    "mAP50-95": 0.457
  },
  "epochs": 100
}

---

Usage

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict("input_video.mp4", conf=0.25, iou=0.7)
results.show()
results.save("output_video.mp4")

---

Citation

@misc{mehdid2026yolopose,
  title={Custom YOLO Pose Estimation for Sprint Analysis},
  author={Mehdid, Samy Abderraouf},
  year={2026}
}

---

License

This project is licensed under the MIT License.