Resource-Optimized Attan Group Activity Classification Using Long Short-Term Convolutional Neural Networks

This research presents a novel approach to classify and differentiate the various steps of Attan, a cultural dance mainly performed in the region of Pakistan and Afghanistan. Many variations of Attan are characterized by different steps involving regional and cultural influences. The classification of such steps is challenging because of: (i) comprehending the spatial-temporal relationship during the dance, (ii) limited prior work addressing the complexity of Attan, and (iii) unavailability of labeled dataset for classification. To address the classification of Attan, this research creates a dataset of 202 videos from YouTube focusing on four key steps namely: “Body Turn (BT)”, “Hand Clap (HC)”, “Head Turn (HT)”, and “Hands Movement (HM)”. Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models have been applied for the classification process. A combined model, Long-term Recurrent Convolutional Networks (LRCN) is proposed that harness the advantages of both CNN and LSTM models for accurate classification. The proposed model split the video dataset (using a 70/30 distribution) into frames, extracts visual features using CNN to preserve spatial information, and feeds these features into an LSTM for capturing temporal information and perform step prediction. The models are evaluated using the standard parameters of accuracy and loss. The results show that the proposed LRCN model achieves highest accuracy of 86% as compared to the standard CNN (ConvLSTM) model with 75% accuracy. Similarly, the proposed LRCN model shows a validation loss of 0.42 when compared to the ConvLSTM with loss of 0.62. Additionally, the training time of LRCN model is minimal when compared to ConvLSTM, making it ideal for real time use.