Human Activity Recognition using BiLSTM and Comparison with Transformer Model

The categorization of human activities using temporal sequence information is attracting interest in deep learning techniques like Recurrent Neural Networks. RNN technique called Long Short Term Memory works well for classifying time series data as it can effectively manage long term dependency and vanishing gradient issues. In this work, we assess the BiLSTM model performance in identifying human activities and comparing it with transformer model. The results show that while the bidirectional approach slows down with increasing dataset size, it improves recognition quality slightly over the LSTM method. In contrast, the transformer model, while not enhancing the accuracy much, can shorten running times when compared to BiLSTM.