Human-Activity-Recognition_using_CNN-LSTM

Developed and implemented deep learning models for human activity recognition, focusing on spatiotemporal data processing and sequence modeling.

Data Pipeline:

• Preprocessed the UCF50 dataset, including frame extraction, resizing, normalization, and conversion to fixed-length sequences to ensure consistent input for training.

  

Model Development:

• Implemented a ConvLSTM model to capture spatiotemporal dynamics by integrating convolutional and LSTM layers.

  

• Additionally, developed an LRCN model combining CNNs for spatial feature extraction and LSTM layers for temporal sequence modeling.

  

Model Training & Optimization:

• Trained models using categorical cross-entropy loss and Adam optimizer, with early stopping callbacks to prevent overfitting.
• The ConvLSTM model was trained over 50 epochs with recurrent dropout, while the LRCN model was trained over 70 epochs with dropout layers after each convolutional block.

Performance Evaluation on the test set:

• Achieved an accuracy of 82.27% with the ConvLSTM model.

• Achieved an accuracy of 89.49% with the LRCN model.