Research on multitarget tracking and event warning technology method for engineering vehicles and personnel based on improved YOLOv5 and DeepSort algorithms

Aiming at the safety supervision needs of engineering vehicles and construction personnel in complex construction site environments, this paper proposes a multi-target tracking and event warning method based on improved YOLOv5 and DeepSort algorithms. At the detection end, the bidirectional weighted feature pyramid (BiFPN) is constructed to replace the neck network of the original YOLOv5, and a fast normalization fusion strategy is adopted to achieve dynamic interaction of cross-scale features, significantly enhancing the feature expression ability for long-distance small targets. Meanwhile, size-optimized anchor box groups and new detection heads are introduced to improve the detection accuracy of construction personnel. At the tracking end, the DeepSort multi-target tracking algorithm is introduced to achieve stable tracking of target trajectories through appearance feature matching and cascaded matching strategies, and a real-time event warning system is constructed based on the dynamic analysis of the safety area of engineering vehicles. Experimental results show that the method effectively improves the robustness of small target detection and cross-camera tracking in complex scenes while maintaining real-time performance, providing technical support for intelligent safety control in construction sites.