Enhanced Human Action Recognition Using PCA and Polynomial Neural Network Fusion

This paper introduces a novel method for improving Human Action Recognition (HAR) by combining Polynomial Neural Networks (PNNs) with Principal Component Analysis (PCA). Our method employs PCA as a preprocessing step to extract robust features and reduce dimensionality, enhancing computational efficiency. These processed features are then fed into custom PNNs, which excel in capturing the complex, nonlinear patterns inherent in human action videos aiming to address the challenges of intra-class variation and occlusions in video data. Experiments conducted on the benchmark dataset of the Berkeley Multimodal Human Action Dataset (MHAD), demonstrate the effectiveness of the PCA-PNN pipeline. Various model architectures using a custom polynomial layer replacing a normal dense layer, such as Deep Neural Networks (DNNs), Recurrent Neural Networks (RNNs), and Long ShortTerm Memory (LSTM) networks, are evaluated across different time window sizes to analyze temporal context impact. Results indicate that the PCA-PNN approach outperforms other models in terms of accuracy. This research contributes to advancing HAR methodologies, offering a robust and efficient solution applicable to real-world scenarios such as surveillance, healthcare, and human-computer interaction.