DynASL-24: Dynamic American Sign Language (ASL) Recognition Dataset

DynASL-24: Dynamic American Sign Language (ASL) Recognition Dataset Overview This dataset contains landmark coordinates and derived geometric features extracted from 24 dynamic American Sign Language (ASL) signs, using the MediaPipe Holistic framework. It was collected to support the training and evaluation of a real-time dynamic sign recognition system combining MediaPipe Holistic and a Long Short-Term Memory (LSTM) network. Dataset Description - Total Samples: 3,600 video sequences (24 signs x 150 sequences each)- Number of Signs: 24 commonly used dynamic ASL signs- Sequence Length: 25 frames per sample- Features per Frame: 285 (218 raw landmark coordinates + 41 engineered angle features + 26 engineered distance features)- Data Format: NumPy arrays (.npy)- Collection Method: Personally recorded videos processed with the MediaPipe Holistic model Sign Inventory (24 signs) Hello, Thanks, Yes, No, Father, Mother, Fine, Please, Forget, Busy, Happy, Sad, Finish, Go, Help, More, Not, Correct, Wrong, Bus, Bad, Good, School, Book Files Included 1. DynamicSigns_DATA(24Signs__150forEach).zip Raw, per-frame data exactly as generated during data collection, organized as follows: DynamicSigns_DATA(24Signs__150forEach)/ <SignName>/ (24 folders, one per sign) <SequenceNumber>/ (150 folders per sign, numbered 1-150) 0.npy ... 24.npy (25 frames per sequence) Each individual .npy file contains a single frame's feature vector of shape (285,). A reference loading script (load_dataset.py, included in this repository) reproduces the exact procedure used by the authors to reconstruct the full dataset into NumPy arrays ready for model training. 2. X_data.npy Consolidated feature array of shape (3600, 25, 285) (float32), obtained by stacking all sequences and frames from the raw data above. Sample order follows alphabetical sign order, then sequence number (1 to 150) within each sign. 3. y_labels.npy Consolidated label array of shape (3600,), containing the sign name (string) corresponding to each sample in X_data.npy, in the same order. 4. label_map.json Dictionary mapping each sign name (string) to an integer index, provided for convenience. Users may apply their own encoding (integer, one-hot, etc.) as required by their framework. 5. load_dataset.py Reference Python script used by the authors to load the raw per-frame .npy files (see File 1) into consolidated X/y arrays for LSTM model training (Keras/TensorFlow). Provided for full transparency and reproducibility of the original experimental pipeline. Feature Engineering Details Raw Landmark Coordinates (218 features) Extracted using MediaPipe Holistic, combining a reduced set of pose landmarks (arms, shoulders, upper body) and full hand landmarks (left and right hands), with (x, y, z, visibility) for pose landmarks and (x, y, z) for hand landmarks. Angle Features (41 features) Inter-landmark angles computed from triplets of pose and hand landmarks, capturing joint flexion and orientation changes characteristic of each dynamic sign. Angle values are normalized to the range [-1, 1]. Distance Features (26 features) Euclidean distances computed from pairs of pose and hand landmarks, capturing spatial configuration and relative movement between key points. The complete list of landmark triplets/pairs used for angle and distance computation is provided in the associated publications (see Related Works below). Usage Example pythonimport numpy as np Option A: load the consolidated arrays directly X = np.load("X_data.npy") # shape (3600, 25, 285)y = np.load("y_labels.npy") # shape (3600,), sign names as strings print(X.shape, y.shape) Option B: reconstruct from raw per-frame files (see load_dataset.py) Related Works This dataset supports the following publications: 1. Y. Farhan and A. Ait Madi, "Real-time Dynamic Sign Recognition using MediaPipe," 2022 IEEE 3rd International Conference on Electronics, Control, Optimization and Computer Science (ICECOCS), 2022, pp. 2-8, doi: 10.1109/ICECOCS55148.2022.9982822. 2. Y. Farhan, Z. Haimer, and A. Ait Madi, "Dynamic Sign Language Recognition: A Hybrid Approach Combining MediaPipe and LSTM," CommIT (Communication and Information Technology) Journal, [in press]. License This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. Contact - Email: [e-mail]- Affiliation: Advanced Systems Engineering Laboratory (LSIA), National School of Applied Sciences, Ibn Tofail University, Kenitra, Morocco Version History - v1.0 (July 2026): Initial release with 3,600 samples covering 24 dynamic ASL signs. Acknowledgments This dataset was created as part of research on sign language recognition systems to enhance communication accessibility for hearing-impaired individuals.