Advancing human action recognition: wavelet-DTW enhanced deep learning with multi-head attention

This study introduces a novel approach to human action recognition by combining discrete wavelet transform (DWT) for multi-scale feature extraction, dynamic time warping (DTW) for sequence alignment, and multi-head attention (MHA) within a convolutional bidirectional long-short-term memory (Conv-BiLSTM) framework. This integration enables precise recognition across diverse temporal scales and complex motion dynamics, setting our approach apart from traditional models. The framework effectively addresses challenges such as imbalanced class distributions and varying action speeds in short video clips, optimising both accuracy and computational efficiency. On benchmark datasets UCF101 and HMDB51, the model achieves 97.02% and 91.6% accuracy, respectively, outperforming current state-of-the-art methods. Statistical analyses and ablation studies demonstrate the contribution of each component to the model's performance. A detailed comparison with other methods highlights its advantages for real-time applications. This work advances human action recognition by combining traditional and modern techniques in an optimised, low-cost architecture suitable for dynamic environments.