A Transfer Learning based approach for Classification of Bhangra Steps

Transfer learning based approaches are increasingly being used for human activity recognition in the last few years, but its scope hasn’t been explored much in the field of dance classification. This paper presents a transfer learning-based approach to classify the five steps of Bhangra namely Dhamal, Jugni, Fasla, Khunda and Lehriya. For this purpose, the keyframes are extracted from videos for each of the five steps. These frames act as references to create the dataset that contained only those images that were similar to keyframes. This dataset is augmented to increase the size of the dataset from 265 images to 1320 images to enhance the model’s generalization performance. The feature extraction is applied, and the extracted features are used to train the model. The model gives a training accuracy of 99.68 percent and a validation accuracy of 96.68 percent.