Attack Behavior Recognition Based on Improved 3D-CNN and ST-GCN Fusion
Dongyang Pan, Shangyou Zeng · 2025
The existing 3D Convolutional Neural Networks (3D-CNNs) suffer from a large number of parameters and high training complexity, while Spatial-Temporal Graph Convolutional Networks (ST-GCNs) are susceptible to sample and label noise when dealing with multi-object videos and lack attention to global video information. To address these issues, this paper proposes an improved fusion method for attack behavior recognition based on 3D-CNN and ST-GCN. The introduction of the separated 3D-CNN effectively reduces the model parameters and training time. Additionally, the model's focus on dynamic information is enhanced through differential input methods, and the use of a Multi-Layer Perceptron (MLP) achieves adaptive feature fusion, further improving the model's accuracy.