Research on a Real-Time Warning System for Unsafe Behaviors in Hydraulic Construction Based on DeepSORT and Improved YOLOv5s

The construction environment of hydraulic engineering is complex, while traditional safety monitoring methods suffer from low efficiency and delayed response. Although static recognition models based on improved YOLOv5s have enhanced detection accuracy, they still cannot assess behavioral persistence and struggle to achieve proactive early warning. To address this, this study integrates the improved YOLOv5s with the DeepSORT algorithm to construct an integrated real-time "detection-tracking-warning" system. The system utilizes DeepSORT to achieve stable personnel tracking in complex scenarios and triggers dynamic warnings based on spatiotemporal behavioral logic. A desktop prototype system was developed using PyQt5/PySide6. Experimental results show that the system achieves a Multiple Object Tracking Accuracy (MOTA) of 86.2% in multi-object occlusion scenarios; the accuracy of unsafe behavior warning exceeds 95%, with an average delay of less than 1.5 seconds. This research accomplishes a transition from passive recognition to proactive warning, providing an intelligent solution for safety management in hydraulic construction.