Human Activity Recognition (HAR) is pivotal in various domains, including entertainment, security, and healthcare. Conventional methods often exhibit limitations: hierarchical spatial feature extractors capture local spatial structures but struggle with long-term dependencies, temporal dependency modeling units effectively learn sequential patterns but may lose fine-grained spatial details, and multi-perspective sequential attention modules selectively emphasize critical temporal features yet can overlook subtle local variations. To address these challenges, we propose an advanced framework that synergistically integrates these three components, effectively mitigating spatial constraints and enhancing temporal sensitivity. First, the hierarchical spatial feature extractor autonomously distills multi-level spatial representations from raw skeletal data, ensuring robust spatial encoding. Next, the temporal dependency modeling unit captures long-range temporal correlations, preserving essential motion dynamics across time. Finally, the multi-perspective sequential attention module adaptively assigns significance to different time steps, allowing the model to focus on the most informative elements while suppressing redundant information. Extensive experiments on the AIR-Act2Act dataset demonstrate the superiority of the proposed framework, achieving 99.40% accuracy on dataset 1 and 98.72% on dataset 2, significantly surpassing traditional spatial (92.02%) and temporal models (96.40%) as well as other state-of-the-art approaches (98.0%).