Robotic Vision-Driven Behavior Recognition Model Based on Adaptive Attention Mechanism and Scene Significance Region Extraction

The capacity to recognize behavior in complex scenes poses a significant challenge within the domains of computer vision and robotic vision. Conventional approaches frequently encounter difficulties in effectively addressing issues such as background interference, perspective variations and intro-class differences, which are prevalent in real-world robotic applications. The proposed intelligent behavior recognition model is based on an adaptive attention mechanism and scene Significance region extraction. It has been specifically designed for robotic perception systems. By simulating the selective attention mechanism of the human visual system, the system achieves precise focusing on key behavior regions. The approach initially employs multi-scale significance detection to localize critical regions within video sequences. The integration of an adaptive spatio-temporal attention mechanism is instrumental in accentuating motion-relevant areas in terms of spatial distribution and prioritizing key frames in terms of temporal sequence. Finally, graph neural networks (GNNs) have been developed to model spatio-temporal relationships between regions for the purpose of performing behavior recognition. The experimental findings, drawn from datasets such as UCF101 and HMDB51, demonstrate the efficacy of this approach in achieving high recognition accuracy and robustness in complex scenes, with an average recognition rate of 94.2%, thereby signifying a substantial improvement over conventional methods. In addition, the validation of the model on a custom robot interaction dataset has been demonstrated to confirm its effectiveness in scenarios such as human–robot collaboration and service robot task understanding. This study achieves superior performance over TimeSformer, ST-GCN and ViT-based frameworks in behavior recognition tasks by integrating multi-scale scene Significance region extraction, adaptive spatio-temporal attention mechanisms and GNN relational modeling. It significantly enhances accuracy and robustness in complex scenes. The innovation lies in simulating human visual selective attention to dynamically focus on key behavior regions, with experiments validating its effectiveness against background interference and viewpoint variations. This suggests that the model has the potential to enhance robotic situational awareness.