A YOLO Architecture Optimized with Deformable Sparse Convolution and Context-Aware Attention

This study develops a real-time safety monitoring system based on improved deformable convolution. This research presents an enhanced YOLO architecture featuring: (1)a C3K2 deformable convolution module with C2PSA attention mechanism for efficient feature learning, and(2)a domain-adapted dataset (Tourist_behavior) integrating annotated instances with six customized safety-critical behavior categories(crowd gathering, illegal parking, etc.). The optimized model demonstrates practical deployment capabilities on edge devices while maintaining detection accuracy. Experimental results demonstrate the enhanced model achieves 87.5% mAP@ 0.5 on Tourist_behavior dataset, surpassing YOLOv9c (85.2 %) and YOLOv8s (84.7 %) in accuracy, while achieving 39 FPS real-time throughput on Jetson Xavier NX under occlusion scenarios. The solution provides an effective technical framework for scenic area safety management while maintaining pose adaptation.