Human Action Recognition Using Log-Softplus Error Activation Function with Bi-Directional Long Short-Term Memory

Human Action Recognition (HAR) involves the recognition of human activities using sensor data. However, existing methods face difficulty with varying viewpoints and occlusion, which reduces accuracy when body parts are hidden or viewed from different angles, making actions harder to interpret effectively. To address these issues, a Log-Softplus Error Activation Function (Serf) with a Bi-Directional Long Short-Term Memory (BiLSTM) network is proposed for accurate recognition of HAR. The preprocessing technique, namely $\mathbf{Z}$-score normalization, ensures normalized input data with unit variance and zero mean, improving training stability. The Synchronized Spatial-Temporal Extraction (SSTE) feature extraction method is used for efficient capture of motion dynamics across space and time. The NTU-RGBD dataset, consisting of complex human activities under varied subjects and viewpoints, is utilized for model evaluation, thereby enhancing the performance of the model. The BiLSTM model, improved with Serf, processes bidirectional temporal dependencies, enhancing learning capacity. The proposed system attains an impressive accuracy of 97.06%, outperforming other existing algorithms such as the GlobalLocal Motion Fusion Network (GLMFN) and Ternary Adaptive Graph Convolution (TAGC).