ProjectKitty

---

🔧 Setup Instructions

Before running any script, create a virtual environment and install the required packages.

1. Create and activate the virtual environment

python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

2. Install dependencies

pip install -r requirements.txt

---

🧠 Script Descriptions

File	Description
baseline_curve.py	Defines baseline motion curves (e.g., joint angles) for comparison during corrective feedback.
baseline_good_curve.json	Stores baseline feature curves of ideal bowling actions. Used for comparison.
bowling_lstm_classifier.keras	Trained LSTM model for classifying good/bad bowling actions.
check_keypoints.py	Validates extracted keypoints from pose estimation; checks for completeness or anomalies.
convert_json.py	Converts raw JSON pose data to the internal standardized format used by the analysis pipeline.
custom_layer.py	Contains custom Keras layers or attention modules used in the LSTM model.
detect_runups.py	Identifies bowling run-up segments in slowed/stabilized videos using pose or movement heuristics.
extract_features.py	Converts keypoints to feature vectors such as joint angles, velocities, or body alignment metrics.
extract_keypoints.py	Runs MediaPipe Pose (optionally after YOLO bowler detection) and outputs structured keypoints in JSON.
launch.py	Unified script to run the complete pipeline: from video input → pose → features → classification → feedback.
lstm_train.py	Trains the LSTM model using the extracted pose features and labeled sequences.
m3u8_downloader.py	Downloads cricket videos (e.g., match highlights) from .m3u8 links. Useful for dataset creation.
slowmo_standardise.py	Slows down and stabilizes cricket videos to standard frame rate for consistent pose detection.
workable.py	Full pipeline script to analyse input videos and give predictions with corrective feedback.
firebase_pose_features2.csv	Feature data collected from Firebase or live upload for training or analysis.
bowling_pose_dataset.csv	Master dataset containing pose features and labels for LSTM training.
captured_video.json	Raw keypoints data from a recorded video.
captured_video_processed.json	Cleaned and standardized version of the above, ready for model input.
bowling_sequences.json	Bowling video segments with annotations (start/end frames).
good_bowling_sequences.json	Subset of sequences labeled as good bowling actions.
training_history.json	Stores model training metrics such as accuracy and loss per epoch.
roboflow.ipynb	Jupyter notebook used for preparing object detection datasets via Roboflow (for YOLO).

---

📁 Folder Summary

Folder	Description
runs/	Likely used for storing YOLOv8/YOLOv11 inference results.
slowed_videos/	Contains slowed and stabilized versions of bowling videos.
output/	Final analysis outputs such as classification results, visualizations, or feedback videos.
launch_vids/	Raw videos captured via the launch.py script.
launch_processed/	Processed video clips generated after trimming and standardisation, these will be sent as input to workable.py.
data/	Raw data that we started with , this was later processed and trimmed and used to train the model.

---

🎥 Pretrained Models

File	Description
yolov8n.pt	Lightweight YOLOv8 model for real-time person/bowler detection.
yolo11n.pt	Custom YOLO model possibly trained for better bowler segmentation or arrow detection.

---

📹 Sample Video

File	Description
detected_with_pose.mp4	Sample video output with overlaid pose keypoints from MediaPipe. Useful for validation.

---

🧪 Workflow Summary

#Real-Time

1. Run launch.py
2. It will capture your video (s to start, 3 sec timer before start, esc to end) please try to capture only the runup part.
3. launch.py will automatically standardise your raw input and call workable.py for analyis and feedback with prediction.
4. If bad action is detected, compare against baseline_good_curve.json for feedback.

#Mp4s

1. Run workable.py with a processed runup clip as sys.argv[1].
2. You can use the samples clips directly or process your own mp4 for input before using it. (Trimmed raw bowling clip -> slowmo_standardise.py -> Ready for workable.py)

---