An Enhanced Approach on Temporal Graph Neural Networks for Human Activity Recognition
V Velantina, V Manikandan, P Manikandan · 2024
Human Activity Recognition systems are valued in a difference of applications, including interactive gaming, surveillance, and healthcare monitoring. This research designates a unique technique for growing the accuracy and effectiveness of HAR systems by combining Temporal Convolutional Networks (TCNs) with Motion History Depth Maps (MHDMs). MHDMs efficiently combine the spatial and temporal dynamics of human motions into single understandable visual demonstration, capturing the essence of motion throughout time. Provisionally, TCNs provide a strong framework for capturing temporal suggestions and subtleties in sequential data while avoiding the problematic and training challenging connected with recurrent neural network designs. We utilized a TCN model to interpret the motion encoded information supplied by MHDMs, enhancing convolutional learnings inherent benefits to manage long-range dependencies and fluctuations in activity duration. The TCNs strategy, which includes dilated convolutions and remaining suggestions, lets a more in-depth analysis of the sequential depth data, guaranteeing that both local and global motion aspects are definitely captured. To test our technique, we ran extensive experiments on benchmark datasets that are widely used in HAR research. Our findings that concerning MHDMs with TCNs provocatively surpasses representative methods in terms of accuracy and computing efficiency. We also examined the influence of numerous in terms on performance, gaining insights into the best settings for dissimilar kinds of activity.