DSGCN: dual-stream graph convolutional network for skeleton-based action recognition under noise interference

The advancement in action recognition is crucial for enhancing interactive systems, improving surveillance accuracy, and optimizing autonomous driving technologies. Traditional RGB-based recognition methods rely on obtaining complete human body information but often face challenges such as sensitivity to environment conditions, loss of spatial information, and interference from ambient noise. These factors pose significant challenges to the accuracy of dynamic action recognition, especially in uncontrolled environments. To address these issues, our proposed Dual-Stream Graph Convolutional Network (DSGCN) leverages the inherent structures of skeleton data. The data flow is divided into two streams: a joint stream and a bone stream, each of which processed through a graph convolutional network. The joint stream is equipped with Class Activation Maps (CAM) to enhance feature recognition. We assessed the DSGCN model using the NTU RGB + D 60 and NTU RGB + D 120 datasets，comparing it with the Richly Activated Graph Convolutional Network (RAGCN) model. By integrating five types of simulated noise into our test, DSGCN demonstrated a 3% higher overall accuracy than RAGCN. The results indicate that despite noise interference, DSGCN maintains an excellent accuracy, demonstrating its Practicality in real-world scenarios with incomplete data and occlusions.