• Model : Trained YOLOv5s to ONNX
• Datasets : Custom dataset built for barbell recognition instead of COCO (trained with a single class)
• Performance Evaluation: Evaluated Precision, Recall, mAP, etc., using val.py
• Epochs: 20
• Final mAP: 86%

• Applied confidence-based filtering and Non-Maximum Suppression (NMS) -> selects representative detections
• Confidence threshold thresh = 0.3

pip install -r requirements.txt

uvicorn main:app --reload

Upload a video file for processing.

Request Body (multipart/form-data):

• file: Video file (.mp4).

Example curl:

curl -X POST -F "file=@your_video.mp4" [http://127.0.0.1:8000/process_video/](http://127.0.0.1:8000/process_video/)

Response (JSON):

{
  "vertical_accuracy": 95.23,
  "video_url": "/get_video/output_unique_filename.mp4"
}

Download the processed video.

Example URL:

http://127.0.0.1:8000/get_video/output_unique_filename.mp4

• Dataset labeling is required.   
• Potential for overfitting due to an insufficient number of training data.

• Performed direct labeling using LabelImg.   
• Potential improvements: Applying Data augmentation (horizontal flip, color variations, etc.) to secure data diversity and enhance the model's generalization performance.

• Impact on trajectory and vertical accuracy in object unrecognition intervals. (Connects the last recognized position and the newly recognized position with a straight line)

• Handling of object unrecognition intervals and re-detection after prolonged object loss. Setup variables related to interval re-detection:

last_detection_frame = -1  # Frame number when the object was last detected
current_frame = 0  # Current frame number
max_frame_gap = int(fps * 1.5)  # Maximum allowed frame gap (1.5 seconds)