Human Action Recognition (HAR) remains a formidable task in the realms of computer vision and pattern recognition. Existing methods for HAR cope with persistent challenges such as vanishing gradients in deep networks, managing extensive training parameters, and ensuring effective feature selection. To tackle these challenges, a Multi-Layer Attention-Enhanced DenseNet for Efficient HAR (MLAEDenseNet) with a Global Average Pooling (GAP) is proposed in this study. A pretrained DenseNet121 is used for effective feature extraction and reducing the vanishing gradient problem. The incorporation of a two-layer attention procedure enables the MLAEDenseNet model to dynamically concentrate on the most relevant spatial-temporal regions in each frame. This allows better discrimination of actions in videos with complex motion patterns. GAP is employed to significantly reduce network parameters, facilitating computational efficiency without compromising performance. A dense layer is subsequently applied to refine feature representations, enabling robust action recognition. Finally, a softmax layer is utilized for accurate action classification. The proposed architecture demonstrates superior performance compared to conventional methods, showcasing its efficacy in discriminating complex actions in diverse video datasets.